PaperlayerView exact revision on arXiv

Source-linked AI summary

Skin disease diagnosis with deep learning: a review

Hongfeng Li, Yini Pan, Jie Zhao, Li Zhang

arXiv:2011.05627v2eess.IVcs.CVcs.LG

TL;DR

Correct skin-disease diagnosis is challenging because it requires clinicians to integrate multiple visual clues. This review surveys deep-learning methods for skin disease diagnosis and reports that published methods have achieved performance similar to dermatologists.

  • Problem

    Earlier reviews focused on traditional machine learning or specific, outdated diagnosis tasks, while correct diagnosis itself requires integrating multiple visual clues.

    Source: §1 · pp. 2, 4

  • Method

    The paper systematically reviews skin diseases, public datasets, deep-learning concepts and architectures, frameworks, evaluation metrics, and diagnosis applications.

    Source: §1 · pp. 3–4

  • Results

    Published deep-learning methods for skin disease diagnosis achieved performance similar to dermatologists.

    Source: §8 · p. 39

  • Takeaways & Limitations

    Developing and validating systems for new imaging techniques requires substantial research and innovative system development.

    Source: §8 · p. 39

    Prior studies commonly used datasets with limited image counts, while labeling larger skin-disease datasets requires expertise and can be difficult and expensive.

    Source: §8.1.1 · p. 40

Abstract

from arXiv · show

Skin cancer is one of the most threatening diseases worldwide. However, diagnosing skin cancer correctly is challenging. Recently, deep learning algorithms have emerged to achieve excellent performance on various tasks. Particularly, they have been applied to the skin disease diagnosis tasks. In this paper, we present a review on deep learning methods and their applications in skin disease diagnosis. We first present a brief introduction to skin diseases and image acquisition methods in dermatology, and list several publicly available skin datasets for training and testing algorithms. Then, we introduce the conception of deep learning and review popular deep learning architectures. Thereafter, popular deep learning frameworks facilitating the implementation of deep learning algorithms and performance evaluation metrics are presented. As an important part of this article, we then review the literature involving deep learning methods for skin disease diagnosis from several aspects according to the specific tasks. Additionally, we discuss the challenges faced in the area and suggest possible future research directions. The major purpose of this article is to provide a conceptual and systematically review of the recent works on skin disease diagnosis with deep learning. Given the popularity of deep learning, there remains great challenges in the area, as well as opportunities that we can explore in the future.

1. Introduction

This review surveys deep learning applications in skin disease diagnosis, covering core concepts, datasets, architectures, evaluation, and task-specific methods. It also examines persistent challenges and possible future directions amid diagnostic difficulty and limited labeled clinical data.

1. Introduction

  • Skin cancer is common, melanoma is the most fatal skin cancer, and early diagnosis and treatment can improve five-year survival by around 14%.Basal cell carcinoma is the most common skin cancer and, although not usually fatal, burdens health care services.Source: §1 · pp. 1–2
  • 75% to 84% is the reported melanoma diagnostic accuracy for inexperienced specialists using dermoscopy, while diagnosis relies on complex visual clues and varies with expertise.Histopathological examination of a biopsy is the diagnostic gold standard; automated screening could help discriminate cancer from non-cancer and distinguish cancer types.Source: §1 · p. 2
  • Traditional CAD systems using handcrafted features can perform well on limited tasks but often do not generalize across skin diseases, motivating automatic feature learning.Handcrafted features are disease-specific and may not capture the diversity of skin diseases; much earlier feature-learning work focused on dermoscopy or histopathology tasks.Source: §1 · pp. 2–3
  • 60.0% top-1 and 80.3% top-3 accuracy were reported by Esteva et al. for a pretrained CNN skin-disease classifier, significantly outperforming human specialists, though labeled clinical-data scarcity limits wider application.Deep networks learn semantic features from large datasets and can address image variation through multiple layers.Source: §1 · p. 3
  • The review systematically surveys recent deep learning advances across skin diseases, public datasets, architectures, and diagnostic tasks, and discusses field challenges and future directions.It focuses on representative works published before or during 2019 rather than aiming to exhaust the literature.Source: §1 · pp. 3–4

2. Skin disease

The section connects skin structure and UV-related melanocyte biology to skin cancer, distinguishing common keratinocyte carcinomas from less common, more deadly melanoma.

2. Skin disease

  • Skin has epidermal, dermal, and hypodermal layers; melanocytes produce UV-protective melanin, and unusual melanocyte growth causes melanoma.The dermis contains collagen fibers, sensors, receptors, blood vessels, and nerve endings, providing skin elasticity and vigor.Source: §2 · pp. 4–5
  • Excess UV can damage DNA and cause abnormal melanin production and pigmented lesions; abnormal melanocyte growth forms melanoma.Mutation of tumor-suppressor genes can make cell growth uncontrollable and lead to tumor formation.Source: §2 · p. 5
  • More than 4.3 million BCC and 1 million SCC cases are diagnosed annually in the United States; these keratinocyte carcinomas are common, while melanoma is less common but more deadly.Melanoma is an aggressive malignancy of melanocytes; BCC and SCC develop from basal and squamous keratinocytes, respectively.Source: §2 · p. 5
  • Lesion shape, size, color, border, and symmetry inform visual assessment, but suspected skin cancer requires biopsy and histology for explicit diagnosis.Melanin location affects lesion color, and dermal melanin is identified as consequential for melanoma; the four categories are SSM, NM, LMM, and ALM.Source: §2 · p. 5

3. Image acquisition and datasets

The section surveys imaging approaches that reveal skin features at different scales and the public datasets used to train and evaluate diagnosis algorithms. It notes that available skin-disease datasets remain small compared with general computer-vision datasets.

3.1. Image acquisition

  • Dermoscopy is a widely used, non-invasive method that reveals subsurface structures, but its observable scope is limited and diagnostic precision depends on dermatologist experience.The review reports that dermoscopy improved diagnosis performance for malignant cases by 50%.Source: §3.1 · p. 6
  • Other acquisition methods include CLSM, OCT, SIA, ultrasound, MRI, and phone or digital-camera images, each providing different diagnostic information or practical tradeoffs.CLSM sensitivity and specificity were 88% and 71%, respectively, and its equipment costs up to $50,000 to $100,000; OCT sensitivity and specificity range from 79% to 94% and 85% to 96%.Source: §3.1 · pp. 6–7

3.2. Datasets

  • Skin-disease datasets contain clinical, dermoscopy, and pathological images; pathological slides provide a diagnostic gold standard, while clinical images commonly come from mobile cameras [58].Dermoscopy images are captured with high-resolution DSLRs or smartphone attachments, and pathological slides are scanned and digitized.Source: §3.2 · p. 7
  • Dermnet and OLE were combined for universal skin-disease classification, with OLE's 19 classes mapped to Dermnet's top-level categories and cross-dataset testing performed.Dermnet includes more than 23,000 images organized across 23 top-level and over 600 fine-grained disease categories; both source datasets contain watermarks.Source: §3.2 · p. 8
  • Public datasets support varied tasks: ISIC provides dermoscopy images for segmentation, feature detection, and classification, while HAM10000 contains 10,015 images spanning pigmented-lesion diagnostic categories [63].Dermofit provides 1,300 standardized lesion images with expert diagnoses and segmentation masks; MoleMap contains 102,451 paired macro and dermoscopy images across 25 conditions [75].Source: §3.2 · pp. 8–10
  • Despite the variety of public datasets, the review characterizes skin-disease datasets as small in samples and patients compared with general computer-vision datasets containing hundreds of thousands or millions of labeled examples.The listed datasets are not exhaustive.Source: §3.2 · p. 11

4. Deep learning

The section introduces deep learning for skin disease diagnosis and outlines neural networks and CNNs, from their basic structure to key image-processing components.

4. Deep learning

  • Deep learning learns layered representations from raw data, while scarce medical labels make semi-supervised and unsupervised methods relevant.Its success is linked to large datasets, powerful computation, and accessible frameworks; models are tuned using training and validation data, then evaluated on test data.Source: §4 · pp. 11–13

4.1. Neural networks

  • Neural networks compose weighted, biased nonlinear transformations across hidden layers; networks with multiple layers are deep neural networks.The section also describes mini-batch stochastic gradient descent and task-specific loss functions, including binary and categorical cross-entropy.Source: §4.1 · pp. 13–14

4.2. Convolutional neural networks

  • CNNs preserve spatial relationships with sparse connections, using convolution, activation, and pooling layers, often followed by fully connected output layers.They learn image features directly, offering an alternative to handcrafted feature engineering.Source: §4.1 · p. 14, §4.2 · p. 15

4.2.1. Convolutional layers

  • Convolution applies shared filters across input locations to produce feature maps, reducing weights because patterns such as lines can recur across an image.The passage illustrates the operation with a 3×3 window moved across a two-dimensional grid.Source: §4.2.1 · p. 15

4.2.2. Activation layers

  • Nonlinear activation enables networks to approximate nonlinear functions; ReLU is common, with leaky ReLU and PReLU as variants.ReLU is defined as max(0, z).Source: §4.2.2 · p. 16

4.2.3. Pooling layers

  • Pooling downsamples feature maps and supports translation invariance; dropout regularizes through stochastic neuron removal, while batch normalization can speed training.Stride-greater-than-one convolution can also downsample, and removing pooling has been reported to simplify networks without sacrificing performance [108]. CNN design also considers task requirements, preprocessing, and computation budget.Source: §4.2.3 · pp. 16–17

5. Deep learning frameworks

Open-source deep learning frameworks provide high-level building blocks for designing, training, and validating neural networks, making complex models such as CNNs easier to implement.

TensorFlow and Keras

  • TensorFlow supports multiple programming languages and CPU/GPU execution, while Keras offers a simpler high-level API for fast experimentation at the cost of configurability.TensorFlow also supports integration of data such as graphs, SQL tables, and images; Keras can run on top of TensorFlow, CNTK, and Theano.Source: §5 · pp. 17, 20

PyTorch

  • PyTorch uses dynamic computation graphs that can be changed during runtime, with automatic gradient computation and support for pretrained models.Its Autograd package builds graphs from tensors and computes gradients; dynamic graphs are useful when a network's memory needs are unknown.Source: §5 · p. 20

Caffe

  • Caffe offers fast image processing and learning, multiple programming interfaces, and pretrained networks through its Model Zoo, but has weaker support for recurrent networks and language modeling.Its interfaces include C, C++, Python, MATLAB, and command line.Source: §5 · p. 20

Sonnet

  • Sonnet, built on TensorFlow, separates Python objects representing network parts from their computational graphs to simplify complex architectures and reproduce DeepMind research.The passage characterizes Sonnet as a flexible functional abstraction tool.Source: §5 · p. 20

Other frameworks

  • MXNet [130] supports many programming languages and parallel computing across GPUs and machines, while other listed frameworks include Gluon, Chainer, Theano, and ONNX.MXNet supports imperative and symbolic programming styles; the frameworks are described as open source and under active development, and built on CUDA and cuDNN.Source: §5 · pp. 20–21

6. Evaluation metrics

The section distinguishes evaluation metrics by task: segmentation commonly uses overlap-based IoU and pixel-level indices, while classification uses image-level metrics including accuracy, sensitivity, specificity, AUC, and precision.

6.1. Segmentation tasks

  • IoU is the most common segmentation metric, measuring predicted-ground-truth overlap relative to their union; its 0-to-1 scale assigns better performance to higher values.It is also known as the Jaccard Index.Source: §6.1 · p. 21
  • Additional segmentation indices use pixel-level true positives, true negatives, false positives, and false negatives; values above 128 count as positive and below 128 as negative.The section also presents equations for these indices, but their labels and numerators are not legible in the supplied passages.Source: §6.1 · p. 21

6.2. Classification tasks

  • Classification metrics include image-level accuracy, sensitivity, specificity, AUC, and precision; AUC measures discrimination between groups using the true-positive rate relative to the false-positive rate.Accuracy, sensitivity, and specificity are measured at the whole-image level rather than the pixel level.Source: §6.2 · p. 22

7. Skin disease diagnosis with deep learning

Section 7 reviews deep-learning applications in skin disease diagnosis, introducing common preprocessing and augmentation methods before organizing prior work by task.

7. Skin disease diagnosis with deep learning

  • The section surveys existing skin disease diagnosis studies that use deep-learning methods.Its review is framed by the popularity of deep learning and its applications to diagnosis tasks.Source: §7 · p. 22
  • The review first introduces common data preprocessing and augmentation methods used in deep learning.Source: §7 · p. 22
  • Existing literature is organized by task type, with the section’s review taxonomy illustrated in Fig. 5.Source: §7 · p. 22

7.1. Data preprocessing and augmentation

The section describes preprocessing that adapts variable-resolution skin images to network inputs while preserving information, and augmentation that addresses limited training data and overfitting.

7.1.1. Data preprocessing

  • Preprocessing should preserve image aspect ratios to limit distortion or information loss, normalize illumination per image, and remove hair or other irrelevant material.Yu et al. reported that subtracting a uniform mean pixel value decreased deep-network performance; removal methods include thresholding, morphology, and deep learning.Source: §7.1.1 · pp. 22–23

7.1.2. Data augmentation

  • Data augmentation addresses scarce labeled medical images by expanding training data, helping networks learn properties such as rotation and translation invariance.The passage attributes medical-data scarcity to disease rarity, patient privacy, expert-labeling requirements, and high acquisition costs.Source: §7.1.2 · p. 23
  • Augmentation spans geometric and color transformations, image mixing, erasing, feature-space methods, adversarial training, generative models, style transfer, and metalearning.Al-Masni et al. rotated 4,000 dermoscopy images by 0°, 90°, 180°, and 270°, reducing overfitting and improving robustness; Yu et al. used rotations and random pixel shifts and reported significant ISIC improvement.Source: §7.1.2 · p. 23

7.2. Applications of deep learning in skin disease diagnosis

This section reviews deep-learning applications in skin disease diagnosis, covering lesion segmentation and classification as well as multitask models that combine related analysis tasks and clinical information. It also surveys applications in image synthesis, cancer detection, augmentation, and segmentation, including approaches targeting diagnostic tasks and data scarcity.

7.2.1. Skin lesion segmentation

  • Segmentation is challenged by low lesion-to-skin contrast, skin-tone and lesion variation, artifacts, lighting, and lesion proximity; preprocessing can mitigate these factors [139, 140, 32, 60, 141].The review stresses that segmentation reliability affects image analysis and that lesion-boundary detection can be difficult when different tumors are close together.Source: §7.2.1 · p. 24
  • An FCN–LSTM method required no image preprocessing and achieved 0.98 average segmentation accuracy and 0.93 Jaccard index on ISIC; U-net achieved 77% Dice versus 61% for C-Means.The hybrid RNN–CNN method outperformed CNN-only methods in the authors’ comparison.Source: §7.2.1 · pp. 24–25
  • MobileGAN combined 1-D non-bottleneck factorization networks with position and channel attention, achieving 97.61% accuracy with 2.35 million parameters on ISIC [168].Other adversarial approaches included a U-net generator and convolutional discriminator, which achieved 0.97 average segmentation accuracy and 0.94 Dice coefficient on PH2 and ISIC.Source: §7.2.1 · p. 27

7.2.2. Skin disease classification

  • CNNs matched tested dermatologists on two biopsy-proven binary tasks, and a CNN trained only on HAM10000 dermoscopy images matched 145 dermatologists on clinical photographs.The tasks in the first comparison were keratinocyte carcinomas versus benign seborrheic keratoses and malignant melanomas versus benign nevi.Source: §7.2.2 · pp. 30–31
  • Ensemble learning combines model predictions, while transfer learning reuses pretrained knowledge to address scarce labeled skin-disease data.The review reports that ensembles can outperform single models and that transfer learning can support performance on small datasets.Source: §7.2.2 · pp. 31–32
  • Attention-guided classification, GAN-assisted semi-supervised learning (0.424 average precision with 140 labeled images) [75], and active learning using up to 50% of samples broaden classification strategies.The review also reports that combining segmentation with classification can improve classification accuracy and that adversarial domain adaptation can address domain shift.Source: §7.2.2 · pp. 33–34

7.2.3. Multi-task learning for skin disease diagnosis

  • Multitask learning [208] shares representations across related tasks and can improve task-specific learning efficiency and prediction accuracy compared with training tasks separately.The review describes MTL as solving multiple learning tasks simultaneously while exploring their commonalities and differences.Source: §7.2.3 · p. 37
  • Multi-task CNNs jointly perform lesion segmentation and classification [210], while body-site and multimodal checklist tasks support lesion diagnosis, retrieval, and localization.Jointly learned body-location information boosted lesion-classification performance; the multimodal model classified 7-point checklist criteria and skin conditions.Source: §7.2.3 · pp. 37–38

7.2.4. Miscellany

  • GANs synthesize realistic high-resolution skin images for data-scarcity mitigation and support augmentation, denoising, and evaluation of clinically meaningful information [212] [213] [218].Reported examples include images up to 1024×512, synthetic images that experts found hard to distinguish from real ones, and GAN comparisons using 2,000 training samples at 256×256 resolution; visual ratings differed in realism, diversity, and artifacts.Source: §7.2.4 · p. 38
  • 0.831 F1 score and 0.675 Youden index exceeded nondermatologic physicians, while algorithm accuracy was comparable with dermatologists for facial keratinocytic cancer detection.The method used R-CNN to generate candidate lesions from clinical photographs, then trained CNNs on labeled image crops.Source: §7.2.4 · p. 38
  • Shades of gray color constancy normalized dermoscopy images while retaining estimated illuminants, then used sampled training-set colors for augmentation in ISIC segmentation and classification experiments [220].Source: §7.2.4 · pp. 38–39
  • 77.8% Jaccard Index and 14% tumour disturb pattern were obtained by a convolutional-recurrent network for hair segmentation with weakly labeled data on ISIC.Encoded features fed recurrent layers to model spatial dependencies between disjointed patches.Source: §7.2.4 · p. 39

8. Discussion

The discussion frames deep-learning skin diagnosis as promising, with reported performance similar to dermatologists, but emphasizes unresolved barriers and the absence of a complete decision-support system.

8. Discussion

  • Deep-learning methods for skin disease diagnosis have made promising progress, with reported performance similar to dermatologists.The authors note that developing and validating algorithms or systems to support new imaging techniques requires substantial research and innovative system development.Source: §8 · p. 39
  • Dermoscopic assessment is subjective and varies with clinician experience, so biopsy may follow; biopsying benign lesions can increase patient anxiety and healthcare costs.Training, time, and experience required to use available and upcoming techniques also pose barriers to early and accurate diagnosis.Source: §8 · p. 39
  • Although automated skin-diagnosis methods exist, no complete decision-support system has been developed; the discussion examines fundamental challenges and suggests ways to address them.The discussion focuses on root causes rather than specific cases encountered.Source: §8 · p. 39

8.1. Challenges

Although deep learning has achieved promising performance in skin disease diagnosis, several issues remain before it can be extensively applied in clinical settings.

8.1.1. Limited labeled skin disease data

  • Larger labeled datasets are needed to prevent overfitting, but labeling is costly, making methods that learn from less labeled data important.The largest publicly available dataset identified in the review, ISIC, contains more than 20,000 skin images.Source: §8.1.1 · p. 40

8.1.2. Imbalanced skin disease datasets

  • Class imbalance, often caused by few positive samples among many benign or negative cases, can bias model results despite weighted-loss training.Obtaining balanced datasets may be as difficult as developing large-scale datasets.Source: §8.1.2 · p. 40

8.1.3. Noisy data obtained from heterogeneous sources

  • Models trained on high-quality dermoscopy images may perform worse on noisy, low-resolution images captured with different equipment, lighting, or distances.The review notes that deep learning algorithms are highly sensitive to images from different equipment.Source: §8.1.3 · p. 40

8.1.4. Lack of diversity among cases in existing skin disease datasets

  • Datasets underrepresent dark-skinned people and rare cancers, risking misdiagnosis and motivating more diverse training data.Models trained on fair-skinned populations may fail on dark skin; datasets may omit diseases such as Merkel cell carcinoma and cutaneous lymphoma.Source: §8.1.4 · pp. 40–41

8.1.5. Missing of medical history and clinical meta-data of patients

  • Most deep-learning studies use images alone, despite the diagnostic value of patient history and clinical metadata, which public datasets often lack.Relevant metadata include age, sex, ethnicity, lesion site, size, and structure; additional clinical information improved dermatologist performance in work.Source: §8.1.5 · p. 41

8.1.6. Explainability of deep learning methods

  • Deep networks’ opaque, complex, and variable learned features make their decisions difficult to interpret, prompting calls for explainability in clinical diagnosis.Refitting the same network with different initializations can yield different intermediate features, and further efforts are needed to address the black-box phenomenon.Source: §8.1.6 · pp. 41–42

8.1.7. Selection of deep neural networks for a specific skin disease diagnosis task

  • Few studies explain how to choose an architecture for a specific diagnosis task, so networks should be designed using disease and data characteristics with domain knowledge.The reviewed work commonly uses popular segmentation or classification architectures and sometimes ensembles multiple networks.Source: §8.1.7 · p. 42

8.2. What can we do next? … 8.2.3. Include additional clinical information to assist skin disease diagnosis

The review proposes directions for addressing challenges in deep-learning skin disease diagnosis. These include expanding and improving skin data and incorporating clinical information into diagnosis models.

8.2. What can we do next?

  • The review draws on skin-disease and related-field literature to suggest future directions for resolving challenges and improving diagnosis performance.Source: §8.2 · p. 42

8.2.1. Obtain massive labeled skin disease data

  • Experienced clinicians can label skin images, while tools such as Fiji, LabelMe and Imagetagger can label large datasets more efficiently.Manual labeling is expensive and time-consuming; the cited tools offer automated or semi-automated alternatives.Source: §8.2.1 · p. 42
  • Integrating existing public skin datasets could create a large-scale training resource, while color constancy methods can normalize lighting in dermoscopy images from heterogeneous sources.The passage also notes that noisy, heterogeneous data can be a problem for model performance.Source: §8.2.1 · p. 42

8.2.2. Increase the diversity of clinical skin data

  • Datasets should cover more cutaneous tumors, normal skin types and dark-skinned populations so models can address broader diagnosis tasks.Most reviewed works involve only limited skin disorders, restricting models to predefined categories such as nevus or melanoma rather than subtypes.Source: §8.2.2 · p. 43

8.2.3. Include additional clinical information to assist skin disease diagnosis

  • Because diagnosis also uses lesion history, individual risk profiles and whole-skin assessment, models should incorporate additional clinical information; a study found that information and close-up images improved performance.Source: §8.2.3 · p. 43
  • Medical documents can be processed with NLP, document analysis and data mining, then combined with skin images in multi-view diagnostic models.The passage also proposes integrating human knowledge into deep-learning algorithms to further improve diagnosis performance.Source: §8.2.3 · p. 43

8.2.4. Fuse handcrafted features with deep networks extracted features · 8.2.5. Employ GANs to synthesize additional data for training deep networks · 8.2.6. Exploit transfer learning and domain adaptation for skin disease diagnosis

The review proposes combining handcrafted and deep features, using GAN-generated images cautiously, and applying transfer learning and domain adaptation to address limited labeled data in skin disease diagnosis.

8.2.4. Fuse handcrafted features with deep networks extracted features

  • Handcrafted features can work well with little labeled data and computation, but their weaker generalization makes them useful mainly as supplements to deep features.The review notes that handcrafted features may underperform features learned by deep networks from massive data.Source: §8.2.4 · p. 43
  • Combining skin lesion elevation, evolution, and geometrical features with deep features may improve performance and reduce the labeled-data requirement.The review also proposes integrating handcrafted feature extraction into deep-network learning; decorrelated color spaces are suggested for border detection.Source: §8.2.4 · pp. 43–44
  • Designing deep networks around skin-disease characteristics and task-specific domain knowledge is proposed as another way to improve performance.Source: §8.2.4 · p. 44

8.2.5. Employ GANs to synthesize additional data for training deep networks

  • GANs can generate realistic synthetic images as additional labeled training data, and several works have applied them to skin disease diagnosis [75].The review presents this approach as potentially helpful when large-scale labeled datasets are unavailable.Source: §8.2.5 · p. 44
  • Because GAN-generated images may differ greatly from the original image distribution, the review suggests pretraining with generated images and fine-tuning on original images.Source: §8.2.5 · p. 44

8.2.6. Exploit transfer learning and domain adaptation for skin disease diagnosis

  • Transfer learning and domain adaptation address limited labeled data; approaches include reusing pretrained semantic features or freezing early network layers and training the remainder.Akhtar et al. used deep-model features for dictionary learning; initial layers may be borrowed because they learn similar filters across diverse images.Source: §8.2.6 · p. 44
  • Prior skin-diagnosis studies used transfer learning or domain adaptation, and developments in other fields such as computer vision may also help.Source: §8.2.6 · pp. 44–45

8.2.7. Develop semi-supervised deep learning methods for skin disease diagnosis · 8.2.8. Explore the possibility of applying reinforce learning for skin disease diagnosis · 8.2.9. Reasonable explanation for predictions produced by deep learning algorithms

The review identifies semi-supervised learning, reinforcement learning, and clinically meaningful explanations as directions for skin disease diagnosis. These address limited labeled data, sequential medical decisions, and the need to explain predictions.

8.2.7. Develop semi-supervised deep learning methods for skin disease diagnosis

  • Semi-supervised learning can leverage abundant unlabeled skin data to mitigate the expense and time required to collect expert-labeled data.Source: §8.2.7 · p. 45
  • A few works [75] have applied semi-supervised learning to skin disease diagnosis, while newer computer-vision models [244] suggest scope for disease-specific methods.Source: §8.2.7 · p. 45

8.2.8. Explore the possibility of applying reinforce learning for skin disease diagnosis

  • Because many medical decisions are sequential and reinforcement learning has shown promise in medical image processing, it may help address skin diagnosis problems.Source: §8.2.8 · p. 45
  • The review reports no known applications of reinforcement learning to skin disease diagnosis, making its use in this area a potential research direction.Source: §8.2.8 · p. 45

8.2.9. Reasonable explanation for predictions produced by deep learning algorithms

  • Explainability hinders clinical use of deep learning, since diagnostic support requires explanations for predictions rather than confidence scores alone.Source: §8.2.9 · p. 45
  • One proposed approach is to explain predictions using the ABCDE criteria or the seven-point skin lesion malignancy checklist [58].Source: §8.2.9 · p. 45

9. Summary

The review systematically surveys recent deep learning applications for skin disease diagnosis, covering the field’s concepts, methods, datasets, evaluation, challenges, and future directions. It describes promising experimental performance while emphasizing that systems require critical testing before clinical use.

9. Summary

  • The review systematically surveys recent deep learning applications in skin disease diagnosis, alongside domain foundations, image acquisition, datasets, architectures, frameworks, evaluation metrics, challenges, and future directions.It aims to provide an intuitive understanding of essential concepts and challenges, and to suggest directions for further work.Source: §9 · p. 46
  • Numerous deep learning systems have achieved comparable or superior diagnosis performance on experimental skin disease datasets, but require critical testing before real-life clinical use.The review highlights the need for reliable automated diagnosis that can be used by experts and non-expert clinicians in diagnostic processes.Source: §9 · p. 46

Related papers

Loading 2011.05627v2…