Deep learning techniques, and namely convolutional neural networks, have become the state-of-the-art approach to solve image classification problems in biomedicineMeijering (2020). These techniques have the potential to revolutionise healthcare Hinton (2018), provided that they become accessible to various populations, a challenge that can be tackled thanks to smartphones or edge devices Marcolino et al. (2018). In this scenario, deep learning models need to be executed on edge devices to reduce latency and preserve privacy. Nevertheless, traditional deep learning models have millions of parameters Tan and Le (2019a) and are known to be expensive in terms of computation, memory, and power consumption. These drawbacks make it challenging to embed them in mobile applications He et al. (2016), a problem that can be faced by using compact deep networks, also known as hardware-aware networks.
Compact neural networks are designed taking into consideration not only the accuracy of the networks, but also their computational complexity. Initially, these networks were manually designed by pruning bigger networksCai et al. (2020), building networks based on operations that are cost-friendly Sandler et al. (2018), or applying a number of network compression techniques Wiedemann and others (2020). Those manual methods have been recently replaced by neural architecture search (NAS) techniques that automatically search for the most accurate and efficient architecture under memory and space constraints Wu et al. (2019). In addition, the quantization of existing standard-size models, a technique that converts a deep learning model’s weights to a lower precision such that it needs less computation, has emerged as an alternative to the construction of new compact models Krishnamoorthi (2018).
In the biomedical context, compact deep networks have been successfully employed for glaucoma detection Li et al. (2020), diabetic retinopathy diagnosis Suriyal et al. (2018) or skin cancer classification Chaturvedi et al. (2020)
. Those models are usually trained by applying transfer learningRazavian et al. (2014)2009), in a new target task) and generally using manually designed compact networks such as ResNet-18 Li et al. (2020) or MobileNet Chaturvedi et al. (2020)
. However, automatically designed and quantized compact models are scarce in this context; probably, due to the fact that they are optimised for natural images from the ImageNet challenge; and, it is not clear whether these models can be properly transferred to biomedical images, or whether they obtain better results than manually designed compact networks. In this work, we try to shed light to those questions by conducting a thorough study of several manually designed, automatically designed and quantized compact networks when working with biomedical datasets.
Another important issue of compact models in the biomedical context is that they usually obtain worse results than standard size networks Sandler et al. (2018); Wu et al. (2019). In this work, we propose to face this drawback by using semi-supervised learning techniques Zhu and Goldberg (2009); a set of techniques that employ both labelled and unlabelled data for training a model. Semi-supervised learning might be especially useful in the biomedical context where annotations are scarce and producing them is a time-consuming task that require expert knowledge Irvin et al. (2019). Semi-supervised learning methods have been employed to improve several models for, among other purposes, classifying chest x-ray abnormalities Ho and Gwak (2020), grading diabetic retinopathy Luo et al. (2020), or analysing gastric x-ray images Li and others (2020). However, semi-supervised learning methods are usually applied with big networks, and their impact on compact networks is not usually explored. Hence, in this paper, we study how compact networks can be combined with semi-supervised learning methods to obtain accurate and efficient models.
Namely, the contributions of this paper are:
We study the performance of several state-of-the-art compact deep networks for image classification with several biomedical datasets. In addition, we compare the performance of these networks with standard size deep models and study whether there is an architecture that outperforms the rest.
We perform a thorough statistical analysis to compare the performance of manually designed, automatically designed, and quantized compact networks to answer whether transfer learning can be successfully employed with them in the biomedical context. Moreover, we analyse if there are significant differences about these 3 families of compact networks.
We explore several semi-supervised learning methods to improve the performance of compact deep networks, and study whether any of those methods is especially suited to work with compact models.
We develop a library that simplifies the construction of compact models using semi-supervised learning methods. This is, up to the best of our knowledge, the first library for applying semi-supervised learning methods to construct deep compact image classification models.
All the experiments, models, and the library for this work are available at https://github.com/adines/SemiCompact.
2 Materials and Methods
In this section, we present the datasets, the compact architectures, the procedures and tools used for training and evaluating those architectures, and the semi-supervised learning methods employed in this work. We start by introducing the datasets that have been used for our experiments.
In this work, we propose a benchmark of 10 partially annotated biomedical datasets, described in Table 1, and evaluate the performance of deep learning models and semi-supervised methods using such a benchmark.
|Dataset||Number of Images||Number of Classes||Description|
|Blindness Kaggle (2019a)||Diabetic retinopathy images|
|Chest X Ray Kermany et al. (2018)||Chest X-Rays images|
|Fungi Arredondo-Santoyo et al. (2019)||Dye decolourisation of fungal strain|
|HAM 10000 Tschandl et al. (2018)||Dermatoscopic images of skin lesions|
|ISIC Codella et al. (2017)||Colour images of skin lesions|
|Kvasir Pogorelov et al. (2017)||Gastrointestinal disease images|
|Open Sprayer Kaggle (2019b)||Dron pictures of broad leaved docks|
|Plants Giselsson et al. (2017)||Colour images of plants|
|Retinal OCT Kermany et al. (2018)||Retinal OCT images|
|Tobacco Kumar et al. (2014)||Document images|
For our study, we have split each of the datasets of the benchmark into two different sets: a training set with the of images and a testing set with the of the images. In addition, for each dataset we have selected images per class using them as labelled images and leaving the rest of the training images as unlabelled images to apply the semi-supervised learning methods. The splits used in our experiments and more information about datasets are available in the project webpage.
2.2 Compact networks
We have explored a variety of manually designed, automatically designed, and quantized architectures. Namely, we have studied 4 manually-designed compact networks (MobileNet Sandler et al. (2018), ResNet-18 He et al. (2016), SqueezeNet Iandola and others (2016), and SuffleNet Zhang and others (2018)), 3 automatically designed compact networks (FBNet Wu et al. (2019), MixNet Tan and Le (2019b), and MnasNet Tan and others (2019)) and 2 quantized compact networks (ResNet-18 quantized and ResNet-50 quantized). In addition, for our experiments, we have considered three standard size networks that are ResNet-50 He et al. (2016), ResNet-101 He et al. (2016), and EfficientNet-B3 Tan and Le (2019a). We provide a comparison of different features of these networks in Table 2.
From Table 2, we can notice that automatically designed networks achieve better results than manually designed architectures when tested on the ImageNet challenge; and quantized networks obtain similar results to their non-quantized counterparts (a 0.1% difference in Top-1 accuracy). In addition, it is also worth noting that the accuracy of compact models is getting closer to bigger models, but using less parameters and requiring less floating-point operations (FLOPs).
|Name||Params (M)||FLOPs (M)||Top-1 acc (%)||Top-5 acc (%)||Design|
2.3 Training and evaluation procedure
All the networks used in our experiments are implemented in PytorchPaszke and others (2019), and have been trained thanks to the functionality of the FastAI library Howard and Gugger (2020) using a GPU Nvidia RTX 2080 Ti with 11 GB RAM. In addition, the two quantized newtorks were built using the Pytorch quantization API 111https://pytorch.org/docs/stable/quantization.html.
In order to train the models, we have used the transfer-learning method presented in Howard and Gugger (2020). This is a two-stage procedure that starts from a model pretrained in the Imagenet challenge Deng and others (2009)
, and can be summarised as follows. In the first stage, we replace the head of the model (that is, the layers that give us the classification of the images), with a new head adapted to the number of classes of each particular dataset. Then, we train these new layers (the rest of the layers stay frozen) with the data of each particular dataset for two epochs. In the second stage, we unfreeze the whole model and retrain all the layers of the model with the new data for fifty epochs. In order to find a suitable learning rate for both the first and second stage, we select the learning rate that decreases the loss to the minimum possible value using the algorithm presented inSmith (2017). Moreover, we employ early stopping based on monitoring the accuracy, and data augmentation Simard et al. (2003) (using flips, rotations, zooms and lighting transformations) to prevent overfitting.
2.4 Semi-supervised methods
Now, we describe the semi-supervised learning methods employed in this work. These methods belong to two families of semi-supervised learning techniques: self-training methods Triguero and others (2015) and consistency regularisation techniques Sajjadi and others (2016).
Self-training is a basic approach that (1) defines a base model that is trained on labelled data, (2) uses the model to predict labels for unlabelled data, and, finally, (3) retrains the model with the most confident predictions produced in (2); thus, enlarging the labelled training set. In a variant of self-training called distillation Hinton and others (2015), a big model is used for (1) and (2), whereas a faster and smaller model than the model trained in (1) is employed in (3). Data and model distillation are also two forms of self-training. In the case of data distillation Huang et al. (2018), given a model trained on manually labelled data, this technique applies such a model to multiple transformations of unlabelled data, ensembles the multiple predictions, and, finally, retrains the model on the union of manually labelled data and automatically labelled data. In the case of model distillation Bucila et al. (2006), several models are employed to obtain predictions of unlabelled data; subsequently, those predictions are ensembled, and used to train a new model. Both techniques can also be combined in a technique called data & model distillation Huang et al. (2018).
In this work, the network used as base model in the plain distillation approach has been ResNet-50 due to the balance between its performance and its efficiency. This network has been also employed in the data distillation method together with 6 different test time augmentation techniques (namely, horizontal flipping, vertical flipping, horizontal and vertical flipping, blurring, Gaussian blurring, and gamma correction). The model distillation method combines the predictions of 3 base networks: ResNet-50, ResNet-101, and EfficientNet-B3. Finally, the model & data distillation method combines the settings of data distillation and model distillation. All the models were trained using the approach presented in the previous section, and a threshold of was used to select robust annotations from the unlabelled data.
The second family of semi-supervised learning methods studied in this work are consistency regularisation techniques Sajjadi and others (2016). This family of methods applies data augmentation to semi-supervised learning based on the idea that a classifier should output the same class distribution for an unlabelled example even after it has augmented. In our case, we have focused on two concrete algorithms: FixMatch Sohn and others (2020) and MixMatch Berthelot and others (2019). As a first step, the FixMatch algorithm generates pseudo-labels using the model’s predictions on weakly-augmented unlabelled images (that is, images that are flipped and translated); subsequently, for a given image, the prediction is only retained if the model produces a high-confidence prediction; and, finally, the model is trained to predict the pseudo-label when feeding a strongly-augmented version of the same image (that is, an image produced using the RandAugment method Cubuk and others (2020)). Similarly, the MixMatch algorithm uses weakly-augmented images both for generating pseudo-labels and for training the model. For the FixMatch and MixMatch algorithms, we have used the default parameters fixed in the two extensions of FastAI that implement them 222https://github.com/oguiza/fastai_extensions 333https://github.com/phanav/fixmatch-fastai.
In order to facilitate the reproducibility of our methods, and also to simplify the application of the aforementioned semi-supervised algorithms to image classification problems, we have designed a Python library, available at the project webpage, that implements them. The library provides an API that is summarised in Figure 1. Several settings can be configured for the methods of the API, and we explain those options in the documentation of the project webpage. In order to employ, for instance, the data distillation method, the user must provide the path to the dataset of images (that should contain labelled and unlabelled images), the name of the base model, the name of the final model (both the compact networks and standard size models studied in this work are supported by the library, and new models can be easily included), the list of transformations that will be applied, and the confidence threshold. From that information, the library will automatically train the base model using the labelled images (the model will be trained following the approach presented in the previous section), use such a base model to annotate the unlabelled images, and train the final model (again using the procedure previously presented). The process is similar, and also automatic, for the rest of the methods provided in the library.
2.5 Statistical analysis
Finally, we explain the evaluation procedure applied to analyse the compact models and semi-supervised learning methods. We have selected the F1-score metric to evaluate the performance of the different networks and methods. In particular, we have studied the F1-score for all binary classification datasets and its macro version for the multi class datasets. The F1-score provides a trade-off between precision and recall, hence we have used this metric to determine whether the results obtained are statistically significant by performing several null hypothesis tests using the methodology presented inGarcia and others (2010); Sheskin (2011)
. In order to choose between a parametric or a non-parametric test to compare the models, we check three conditions: independence, normality, and heteroscedasticity — the use of a parametric test is only appropriate when the three conditions are satisfiedGarcia and others (2010).
The independence condition is fulfilled in our study since the models and methods are tested on independent datasets. We use the Shapiro-Wilk test to check normality — with the null hypothesis being that the data follow a normal distribution — and, a Levene test to check heteroscedasticity — with the null hypothesis being that the results are heteroscedastic. Since more than two models are involved in all the conducted tests, we employ an ANOVA test if the parametric conditions are fulfilled, and a Friedman test otherwise. In both cases, the null hypothesis is that all the models have the same performance. Once the test for checking whether a model is statistically better than the others is conducted, a post-hoc procedure is employed to address the multiple hypothesis testing among the different models. A Bonferroni-Dunn post-hoc procedure, in the parametric case or a Holm post-hoc procedure, in the non-parametric case, is used for detecting significance of the multiple comparisons and thevalues should be corrected and adjusted Garcia and others (2010); Sheskin (2011). We have performed our experimental analysis with a level of confidence equal to . In addition, the size effect has been measured using Cohen’s d Cohen (1969).
3 Results and Discussion
In this section, we present a thorough analysis for the results obtained by the 9 compact networks and the 6 semi-supervised processes when applied to the datasets introduced in the previous section. In particular, we first compare the performance of the networks when training using only the labelled data. Then, we study the impact of applying the different semi-supervised methods. From the results obtained in our experiments, we can answer some questions that have driven this research.
|Blindness||Chest X Ray||Fungi||HAM 10000||ISIC||Kvasir||Open Sprayer||Plants||Retinal OCT||Tobacco||Mean(std)|
Mean (and standard deviation) F1-score for the standard size models (ResNet-50, ResNet-101 and EfficientNet), compact models (FBNet, MixNet, MNasNet, MobileNet, ResNet-18, SqueezeNet, and ShuffleNet) and quantized models (ResNet-18 quantized and ResNet-50 quantized) for the base training method on the 10 biomedical datasets. The best result is highlighted in bold face.
First of all, we have studied which architecture produce the most accurate model for each dataset and whether there are significant differences among them. As we can see in Table 3, ResNet-50 obtains the best results in 4 datasets, MixNet (one of the compact architectures) in 4, ResNet-101 in one, and EfficientNet in another one. In addition, as we show in the first row of Table 4, there are not significant differences among standard size models and 4 compact models that are FBNet, MixNet, MnasNet and ResNet-18.
We have also conducted an analysis of the performance of each architecture trained using the different semi-supervised learning methods — a summary of the results is presented in Table 5. A first conclusion from such an analysis, see Table 4, is that there are not significant differences among the FBNet, MixNet, MnasNet and ResNet-18 compact networks and the three standard size models when we apply any self-training method. In fact, those compact networks outperform the standard size models when the Plain Distillation and Data Distillation training methods are applied. Similarly, compact networks generally outperform standard size models when consistency regularisation methods are employed.
|Test (ANOVA or Friedman)||After post-hoc procedure|
|Base||ResNet-50 EfficientNet, FBNet, MixNet, MNasNet, ResNet-101, ResNet-18; ResNet-50 MobileNet, SqueezeNet, ShuffleNet, ResNet-18 quantized, ResNet-50 quantized;|
|Plain||ResNet-18 EfficientNet, FBNet, MixNet, MNasNet, ResNet-50, ResNet-101; ResNet-18 MobileNet, SqueezeNet, ShuffleNet, ResNet-18 quantized, ResNet-50 quantized;|
|Data||MixNet EfficientNet, FBNet, MNasNet, ResNet-50, ResNet-101, ResNet-18; MixNet MobileNet, SqueezeNet, ShuffleNet, ResNet-18 quantized, ResNet-50 quantized;|
|Model||ResNet-50 EfficientNet, FBNet, MixNet, MobileNet, ResNet-101, ResNet-18, SqueezeNet; ResNet-50 MNasNet, ShuffleNet, ResNet-18 quantized, ResNet-50 quantized;|
|DataModel||ResNet-50 EfficientNet, FBNet, MixNet, MNasNet, MobileNet, ResNet-101, ResNet-18; ResNet-50 SqueezeNet, ShuffleNet, ResNet-18 quantized, ResNet-50 quantized;|
|FixMatch||ResNet-18 MixNet, MNasNet, MobileNet, ResNet-18 quantized, ResNet-50 quantized; ResNet-18 EfficientNet, FBNet, ResNet-50, ResNet-101, SqueezeNet, ShuffleNet;|
|MixMatch||MNasNet EfficientNet, FBNet, MobileNet, ResNet-18, ResNet-50 quantized; MNasNet MixNet, ResNet-50, ResNet-101, SqueezeNet, ShuffleNet, ResNet-18 quantized;|
|ResNet-50||ResNet-101||EfficientNet||FBNet||MixNet||MNasNet||MobileNet||ResNet-18||SqueezeNet||ShuffleNet||ResNet-18 quantized||ResNet-50 quantized|
We have also studied the difference in performance among families of compact networks (manually designed, automatically designed, and quantized). From the results presented in Table 6, we can conclude that automatically designed compact networks obtain the best results with the base training and all the semi-supervised methods, except for FixMatch; and in most cases have a significant difference with manually designed networks. On the contrary, quantized networks, generally have worse performance than NAS and manually designed networks, except when trained with the FixMatch method, that they achieve the best performance.
|NAS||Manual||Quantized||Test (ANOVA or Friedman)||After post-hoc procedure|
|Base||81.0(14.7)||73.4(14.6)||73.9(15.9)||NAS Manual, Quantized;|
|Plain||83.1(13.2)||75.6(15.2)||73.8(17.7)||NAS Manual, Quantized;|
|Data||83.8(12.4)||76.0(15.0)||72.2(18.1)||NAS Manual, Quantized;|
|Model||81.0(12.4)||74.7(14.7)||59.5(27.4)||NAS Manual; NAS Quantized;|
|DataModel||80.5(14.0)||75.0(15.1)||57.4(25.8)||NAS Manual; NAS Quantized;|
|FixMatch||70.5(21.8)||65.3(17.0)||76.4(19.8)||Quantized NAS; Quantized Manual;|
|MixMatch||66.0(17.8)||58.4(15.3)||63.1(21.5)||NAS Manual, Quantized;|
In addition, we have also explored which semi-supervised learning method obtains the best results for each network, see Tables 7 and 8. For the three standard size models, the self-training methods obtain a similar result to the base training approach and outperform the consistency regularisation methods. In particular, the Data Model Distillation approach achieves the best results and has a significant difference with the FixMatch and MixMatch methods, but not with the base training method. For the automatically designed compact networks, the Data Distillation method outperforms the rest; and, as in the previous case, there are not significant differences with the base training method. For manually designed compact networks, there is not a single method that stands out, however, there is a significant difference between the semi-supervised winner method and the base method. Finally, for quantized networks, no method outperforms the rest.
|Test (ANOVA or Friedman)||After post-hoc procedure|
|ResNet-50||DataModel Base, Plain, Data, Model; DataModel FixMatch, MixMatch;|
|ResNet-101||DataModel Base, Plain, Data, Model; DataModel FixMatch, MixMatch;|
|EfficientNet||DataModel Base, Plain, Data, Model; DataModel FixMatch, MixMatch;|
|FBNet||Data Base, Plain, Model, DataModel; Data FixMatch, MixMatch;|
|MixNet||Data Base, Plain; Data Model, DataModel, FixMatch, MixMatch;|
|MNasNet||Data Base, Plain, DataModel, FixMatch, MixMatch; Data Model;|
|MobileNet||Model Plain, Data, DataModel, FixMatch, MixMatch; Model Base;|
|ResNet-18||Plain Data, Model, DataModel; Plain Base, MixMatch;|
|SqueezeNet||Plain Data, Model, DataModel; Plain Base, FixMatch, MixMatch;|
|ShuffleNet||Base Plain, Data, Model, DataModel, MixMatch; Base FixMatch;|
|ResNet-18 quantized||Data Base, Plain, Model, FixMatch; Data DataModel, MixMatch;|
|ResNet-50 quantized||FixMatch Data; FixMatch Base, Plain, Model, DataModel, MixMatch;|
In our next set of experiments, we analysed whether there is any semi-supervised method that has a better performance than the rest, and, similarly, which semi-supervised method produces the greatest improvement over the base training approach. In Tables 9 and 11, we can see that the Data Distillation method obtains the best value in the Friedman’s test average ranking. In addition, from the results presented in Tables 10 and 12, we can notice that there is a significant difference in the performance between the Data Distillation method and the rest, except for the Plain Distillation method.
|Method||F1-Score||Friedman’s test average ranking|
|Method||Z value||p value||adjusted p value||Cohen’s d|
|Method||F1-Score||Friedman’s test average ranking|
|Method||Z value||p value||adjusted p value||Cohen’s d|
Finally, we have studied the efficiency of the different networks, see Table 13. In particular, we have analysed the model size, the time that takes each model to complete a training epoch, and, the time that takes each model when applied for inference with a given image. From the results presented in Table 13, we can see that there is a great difference in size between the standard size and compact models. Standard size models are all above 100 MB, whereas the majority of compact models are under 100 MB (being ResNet-18 the only exception). In particular, the difference in size between compact and standard size networks ranges from to , standing out the ShuffleNet network, however the accuracy of this model is usually lower than the rest. We can also notice that the quantized models reduce the size of original models by almost a . Another important point when we test the efficiency of a model is the training time per epoch. In our experiments, using a Nvidia RTX 2080 Ti GPU with 11 GB RAM, standard size networks took approximately between 3 and 4 minutes per epoch. On the contrary, compact networks took less than 100 seconds per epoch (being MixNet the only exception). In this aspect, ShuffleNet stands out from the rest, taking only 15 seconds per epoch. Furthermore, quantized networks reduce the training time per epoch around a from its standard version. Finally, we have studied the inference time of each model. We have calculated the time that takes each model to infer the class of an image in a Intel(R) Xeon(R) W-2145 CPU with 3.70GHz, 16 CPUs cores and 32 GB. Here, we notice again the difference between standard size and compact models. While standard size models take between 200 and 300 ms to classify an image, the compact networks take, at maximum, half the time (the exception is MixNet that takes almost the same time than standard models). This is even more prominent in the case of ResNet-18, SqueezeNet, ShuffleNet, ResNet-18 quantized, and ResNet-50 quantized, that take, at least 4 less time than standard size networks. In particular, quantized versions of the models take between and less time than their standard size version.
|Network||Size (MB)||Train Time (s)||Inference Time (ms)|
4 Conclusions and further work
In this paper, we have explored how to combine compact networks with semi-supervised learning models. The results show that, with this combination, we can create compact models that are not only as accurate as bigger models, but also faster and lighter. In particular, we have noticed that when training a model without using semi-supervised methods there are compact networks (FBNet, MixNet, MNasNet and ResNet-18) that obtain a similar performance to standard size networks. Also, when we apply semi-supervised learning methods, specifically, the Plain Distillation method and the Data Distillation method, those models outperform the standard size models. In particular, the best results are obtained when we apply Data Distillation to MixNet and Plain Distillation to ResNet-18. Another conclusion that we can draw from our study is that, in general, automatically designed networks obtain better results than manually designed networks and quantized networks, with the exception of ResNet-18 that obtained similar results than automatically designed networks. Regarding the question of which semi-supervised method produce the best results for each king of architecture, we can conclude that the Data Model Distillation method is the best option for standard size networks; that Data Distillation is the best for automatically designed networks; and, in the rest of cases, there is no a general rule, although the Data Distillation approach generally obtains good results. Finally, compact networks outperform standard size networks in efficiency, that is, in size and in speed. In summary, with this study we can conclude that by applying the Data Distillation method to MixNet, or Plain Distillation to ResNet-18 we can obtain models as accurate as standard size model but, also, faster and lighter. Since these rules might not be the best alternative for a particular case, we have developed a library that simplifies the process of training compact image classification models using semi-supervised learning methods.
In the future, we plan to extend our work to object detection and semantic segmentation tasks, two instrumental tasks in biomedical images. In those problems, deep models employ as backbones architectures like ResNet or EfficientNet, and we plan to study the impact of replacing those backbones with compact models. Another task that remains as further work is the development of a simple approach to integrate compact models into smartphones and edge devices. This is instrumental to facilitate the creation and dissemination of deep learning models for biomedicine.
This work was partially supported by Ministerio de Ciencia e Innovación [PID2020-115225RB-I00 / AEI / 10.13039/501100011033], and FPU Grant 16/06903 of the Spanish MEC.
Automatic characterisation of dye decolourisation in fungal strains using expert, traditional, and deep features. Soft Computing 23, pp. 12799–12812. External Links: Cited by: Table 1.
- Mixmatch: a holistic approach to semi-supervised learning. In 33rd International Conference on Neural Information Processing Systems (NEURIPS’19), pp. 5050–5060. Cited by: §2.4.
- Model compression: making big, slow models practical. In 12th International Conference on Knowledge Discovery and Data Mining, KDD’06, pp. 535–541. Cited by: §2.4.
- Once for all: train one network and specialize it for efficient deployment. In International Conference on Learning Representations, External Links: Cited by: §1.
Skin lesion analyser: an efficient seven-way multi-class skin cancer classification using mobilenet.
AMLTA 2020: Advanced Machine Learning Technologies and Applications, Advances in Intelligent Systems and Computing, Vol. 1141, pp. 165–176. External Links: Cited by: §1.
- Skin Lesion Analysis toward Melanoma Detection: A Challenge at the International Symposium on Biomedical Imaging (ISBI) 2016. In , CVPR’17. Cited by: Table 1.
- Statistical power analysis for the behavioral sciences. Academic Press, USA. Cited by: §2.5.
- Randaugment: practical automated data augmentation with a reduced search space. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pp. 3008–3017. External Links: Cited by: §2.4.
- ImageNet: A Large-Scale Hierarchical Image Database. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR09), Cited by: §1, §2.3.
- Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: experimental analysis of power. Information Sciences 180, pp. 2044–2064. Cited by: §2.5, §2.5.
- A Public Image Database for Benchmark of Plant Seedling Classification Algorithms. arXiv preprint abs/1711.05458. Cited by: Table 1.
- Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778. External Links: Cited by: §1, §2.2.
- Distilling the knowledge in a neural network. External Links: Cited by: §2.4.
- Deep learning – a technology with the potential to transform health care. JAMA 320 (11), pp. 1101–1102. External Links: Cited by: §1.
- Utilizing knowledge distillation in deep learning for classification of chest x-ray abnormalities. IEEE Access 8, pp. 160749–160761. External Links: Cited by: §1.
- Fastai: a layered api for deep learning. Information 11, pp. 108. External Links: Cited by: §2.3, §2.3.
- Omni-supervised learning: scaling up to large unlabelled medical datasets. In Medical Image Computing and Computer Assisted Intervention, MICCAI’18, pp. 572–580. Cited by: §2.4.
- SqueezeNet: alexnet-level accuracy with 50x fewer parameters and 0.5mb model size. External Links: Cited by: §2.2.
CheXpert: a large chest radiograph dataset with uncertainty labels and expert comparison.
Thirty-Third AAAI Conference on Artificial Intelligence, AAAI’19, Vol. 33, pp. 590–597. Cited by: §1.
- APTOS 2019 blindness detection. Note: https://www.kaggle.com/c/aptos2019-blindness-detection Cited by: Table 1.
- Open sprayer images. Note: $https://www.kaggle.com/gavinarmstrong/open-sprayer-images$ Cited by: Table 1.
- Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell 172 (5), pp. 1122 – 1131.e9. External Links: Cited by: Table 1.
- Quantizing deep convolutional networks for efficient inference: a whitepaper. arXiv preprint arXiv:1806.08342. Cited by: §1.
- Structural similarity for document image classification and retrieval. Pattern Recognition Letters 43, pp. 119 – 126. Note: ICPR2012 Awarded Papers External Links: Cited by: Table 1.
- Development and clinical deployment of a smartphone-based visual field deep learning system for glaucoma detection. npj Digital Medicine 3, pp. 120. External Links: Cited by: §1.
- Soft-label anonymous gastric x-ray image distillation. In 2020 IEEE International Conference on Image Processing (ICIP), pp. 305–309. External Links: Cited by: §1.
- Automatic diabetic retinopathy grading via self-knowledge distillation. Electronics 9 (9), pp. 1337. External Links: Cited by: §1.
- The impact of mhealth interventions: systematic review of systematic reviews. JMIR Mhealth and Uhealth 6 (1), pp. e23. External Links: Cited by: §1.
- A bird’s-eye view of deep learning in bioimage analysis. Computational and Structural Biotechnology Journal 18 (1), pp. 2312–2325. External Links: Cited by: §1.
- PyTorch: an imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett (Eds.), pp. 8024–8035. Cited by: §2.3.
- KVASIR: a multi-class image dataset for computer aided gastrointestinal disease detection. In 8th ACM on Multimedia Systems Conference, MMSys’17, pp. 164–169. External Links: Cited by: Table 1.
- CNN features off-the-shelf: An astounding baseline for recognition. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, CVPRW’14, pp. 512–519. Cited by: §1.
- Regularization with stochastic transformations and perturbations for deep semi-supervised learning. In 30th International Conference on Neural Information Processing Systems (NIPS’16), pp. 1171–1179. Cited by: §2.4, §2.4.
- MobileNetV2: inverted residuals and linear bottlenecks. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4510–4520. External Links: Cited by: §1, §1, §2.2.
- Handbook of parametric and nonparametric statistical procedures. CRC Press, London. Cited by: §2.5, §2.5.
- Best practices for convolutional neural networks applied to visual document analysis. In Proceedings of the 12th International Conference on Document Analysis and Recognition (ICDAR’03), Vol. 2, pp. 958–964. Cited by: §2.3.
- Cyclical learning rates for training neural networks. In IEEE Winter Conference on Applications of Computer Vision, WACV’17, pp. 464–472. External Links: Cited by: §2.3.
- FixMatch: simplifying semi-supervised learningwith consistency and confidence. In 34th International Conference on Neural Information Processing Systems (NEURIPS’20), Cited by: §2.4.
- Mobile assisted diabetic retinopathy detection using deep neural network. In 2018 Global Medical Engineering Physics Exchanges/Pan American Health Care Exchanges (GMEPE/PAHCE), pp. 1–4. External Links: Cited by: §1.
- EfficientNet: rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, ICML 2019, Vol. 97, pp. 6105–6114. Cited by: §1, §2.2.
- MixConv: mixed depthwise convolutional kernels. In 30th British Machine Vision Conference 2019 (BMVC 2019), pp. 1–13. Cited by: §2.2.
- MnasNet: platform-aware neural architecture search for mobile. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2815–2823. External Links: Cited by: §2.2.
- Self-labeled techniques for semi-supervised learning: taxonomy, software and empirical study. Knowledge and Information Systems 42, pp. 245–284. External Links: Cited by: §2.4.
- The HAM10000 Dataset: A Large Collection of Multi-Source Dermatoscopic Images of Common Pigmented Skin Lesions. Scientific Data 5, pp. . External Links: Cited by: Table 1.
- DeepCABAC: a universal compression algorithm for deep neural networks. IEEE Journal of Selected Topics in Signal Processing 14 (4), pp. 700–714. External Links: Cited by: §1.
- FBNet: hardware-aware efficient convnet design via differentiable neural architecture search. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 10726–10734. External Links: Cited by: §1, §1, §2.2.
- ShuffleNet: an extremely efficient convolutional neural network for mobile devices. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 6848–6856. External Links: Cited by: §2.2.
- Introduction to Semi-Supervised Learning. Synthesis Lectures on Artificial Intelligence and Machine Learning, Morgan & Claypool Publishers. Cited by: §1.