PaperlayerView exact revision on arXiv

Source-linked AI summary

Dermatologist Level Dermoscopy Skin Cancer Classification Using Different Deep Learning Convolutional Neural Networks Algorithms

Amirreza Rezvantalab, Habib Safigholi, Somayeh Karimijeshni

arXiv:1810.10348v1cs.CVcs.LGstat.ML

TL;DR

Skin-lesion diagnosis is challenging, and the paper tests whether deep learning can classify multiple dermoscopy categories against dermatologist performance. Four pre-trained CNNs were evaluated on two public datasets; the models outperformed dermatologists overall, with DenseNet 201 achieving the highest averaged AUC scores.

  • Problem

    Skin-lesion diagnosis is difficult and can be subjective, motivating evaluation of automated classification across multiple diagnostic categories.

    Source: §1 · pp. 1–2

  • Method

    Four ImageNet-pretrained CNNs were evaluated for eight-category classification using combined HAM10000 and PH2 dermoscopy images.

    Source: §2.1 · p. 3

  • Results

    CNNs trained on 10,135 dermoscopy images outperformed dermatologists; melanoma ROC AUC was 94.4% for ResNet 152 versus 82.26% for dermatologists.

    Source: §3 · p. 10, §5 · p. 12

  • Takeaways & Limitations

    The models may be implemented in dermoscopy systems or smartphones to assist dermatologists.

    Source: §5 · p. 12

    The dataset was highly unbalanced, and no preprocessing step was applied.

    Source: §5 · p. 12

Abstract

from arXiv · show

In this paper, the effectiveness and capability of convolutional neural networks have been studied in the classification of 8 skin diseases. Different pre-trained state-of-the-art architectures (DenseNet 201, ResNet 152, Inception v3, InceptionResNet v2) were used and applied on 10135 dermoscopy skin images in total (HAM10000: 10015, PH2: 120). The utilized dataset includes 8 diagnostic categories - melanoma, melanocytic nevi, basal cell carcinoma, benign keratosis, actinic keratosis and intraepithelial carcinoma, dermatofibroma, vascular lesions, and atypical nevi. The aim is to compare the ability of deep learning with the performance of highly trained dermatologists. Overall, the mean results show that all deep learning models outperformed dermatologists (at least 11%). The best ROC AUC values for melanoma and basal cell carcinoma are 94.40% (ResNet 152) and 99.30% (DenseNet 201) versus 82.26% and 88.82% of dermatologists, respectively. Also, DenseNet 201 had the highest macro and micro averaged AUC values for overall classification (98.16%, 98.79%, respectively).

1. Introduction

Skin cancer is widespread and melanoma is especially dangerous, yet skin-lesion diagnosis can be subjective and difficult even for experts. The paper motivates automated, multi-class dermoscopy classification as a way to address gaps in existing methods and support clinicians.

1. Introduction

  • Dermoscopy is a non-invasive way to capture high-resolution skin images that reveal features invisible to the naked eye.The paper presents it as a means of improving visual assessment of lesions.Source: §1 · p. 2
  • Expert skin-lesion diagnosis can be subjective and time-consuming, and reported melanoma sensitivity is often below 80%.Dermoscopy can perform poorly in unskilled hands, and highly trained dermatologists are not available worldwide.Source: §1 · p. 2
  • Earlier image-analysis methods faced insufficient data and limited class diversity, while some relied heavily on hand-made segmentation criteria.The paper also describes traditional machine-learning pipelines as depending on hand-engineered feature representations.Source: §1 · p. 2
  • Deep learning can learn hierarchical representations directly from raw data without expert feature engineering.The paper contrasts this representation learning with earlier approaches that transformed raw inputs into hand-made feature vectors.Source: §1 · pp. 2–3
  • The study tests pre-trained CNNs on eight diagnostic categories from two public dermoscopy databases to compare their accuracy with dermatologists.The stated aim is to support better diagnosis across skin-cancer categories.Source: §1 · p. 3

2. Methods and materials

The study combines two public dermoscopy datasets to classify eight diagnostic categories, then evaluates four ImageNet-pretrained CNNs using training, validation, and test splits. Each model is fine-tuned with a classification head suited to its architecture.

2.1. Materials

  • HAM10000 images and selected PH2 images form an eight-category dataset spanning melanoma, nevi, carcinomas, keratosis, dermatofibroma, and vascular lesions.The study uses 10,015 HAM10000 images and 120 PH2 images: atypical nevi and melanoma.Source: §2.1 · pp. 3–4

2.2. Methods

  • Four ImageNet-pretrained CNNs were evaluated with a 70% training, 15% validation, and 15% testing split.Figure 2 presents the architectures of the four models.Source: §2.2 · p. 4, Figure 2 · p. 5

2.2.1. Google’s Inception v3

  • Inception v3 was fine-tuned across all layers with an average-pooling layer, two fully connected layers, and a softmax output for eight categories.Its inputs were resized to 299 × 299 pixels.Source: §2.2.1 · p. 5

2.2.2. InceptionResNet v2

  • InceptionResNet v2 was fine-tuned across all layers with global average pooling, a fully connected layer, and softmax for eight categories.Its inputs were resized to 224 × 224 pixels.Source: §2.2.2 · p. 5

2.2.3. ResNet 152

  • ResNet 152 was fine-tuned across all layers with average pooling and a fully connected layer before softmax classification of eight categories.Its inputs were resized to 224 × 224 pixels.Source: §2.2.3 · pp. 5–6

2.2.4. DenseNet 201

  • DenseNet 201 was fine-tuned across all layers with global average pooling and softmax to classify eight categories.Its inputs were resized to 224 × 224 pixels.Source: §2.2.4 · p. 6

3. Results

Across melanoma and basal cell carcinoma, all four CNNs achieved higher ROC AUC than dermatologists. DenseNet 201 led the models on overall macro- and micro-averaged ROC AUC, though category-level errors remained.

3. Results

  • Because the dataset is unbalanced, the evaluation reports F1-score and macro- and micro-averaged ROC AUC; ROC curves show the sensitivity-specificity trade-off.F1-score is the harmonic mean of precision and recall.Source: §3 · p. 6
  • 98.16% macro average ROC AUC and 98.79% micro average ROC AUC made DenseNet 201 highest among the models on both measures.The paper reports these averages because the dataset is unbalanced.Source: §3 · p. 6, Figure 3 · p. 7
  • Except for benign keratosis and AKIEC, DenseNet 201 and ResNet 152 achieved the highest ROC AUC scores among the models for individual diagnostic categories.The comparison covers each diagnostic category.Source: §3 · p. 7, Table 1 · p. 8
  • All four models confused melanoma with melanocytic nevi, performed poorly on AKIEC and benign keratosis, while DenseNet 201 and ResNet 152 performed better overall.The confusion matrices’ main diagonal gives recall for each diagnostic category.Source: §3 · p. 8, Figure 4 · p. 9
  • 94.4% ROC AUC for melanoma (ResNet 152) versus dermatologists’ 82.26%, and 99.3% for basal cell carcinoma (DenseNet 201) versus 88.82%; all four CNNs exceeded experts on both.The comparison uses dermatologists’ averaged sensitivity and specificity as the benchmark.Source: §3 · p. 10, Figure 3 · p. 10, Table 3 · p. 11

4. Discussion

Across the eight diagnostic categories, the CNN models outperformed dermatologists on melanoma and basal cell carcinoma, and DenseNet 201 had the strongest overall averaged ROC AUC. The discussion also compares these findings with prior skin-cancer classification studies and describes this work as the first multiclass study using only dermoscopy images.

4. Discussion

  • 94.40% melanoma ROC AUC (ResNet 152) and 99.30% basal cell carcinoma ROC AUC (DenseNet 201) exceeded expert values of 82.26% and 88.82%, respectively.All four trained models exceeded the expert ROC AUC values for both cancer categories.Source: §4 · p. 11
  • 98.79% micro and 98.16% macro ROC AUC: DenseNet 201 led overall, with the best reported micro/macro precision and F1-score also attributed to this model.Micro/macro precision was 89.01%/85.24%, and micro/macro F1-score was 89.01%/85.13%.Source: §4 · p. 11
  • DenseNet 201 led basal cell carcinoma (99.3%) and dermatofibroma (99.9%), while ResNet 152 led melanoma (94.4%) and melanocytic nevi (97.3%).Source: §4 · p. 11
  • Both DenseNet 201 and ResNet 152 reached 100% AUC for atypical nevi and vascular lesions; Inception v3 led AKIEC (98.4%) and benign keratosis (97.1%).Source: §4 · p. 11
  • The authors describe this as the first multiclass study using only dermoscopy images and compare it with prior studies using different datasets and tasks.Examples include VGG-16's 82.6% ROC AUC for melanoma on ISBI 2016 and Inception v4's 86% versus dermatologists' 79% and 82%.Source: §4 · p. 12

5. Conclusion and Future works

The study reports that deep CNNs trained on 10,135 dermoscopy images outperformed dermatologists in classifying eight skin diseases. The authors note dataset limitations and suggest possible clinical-assistance applications and directions for further improvement.

5. Conclusion and Future works

  • Deep CNNs trained on 10,135 dermoscopy images outperformed dermatologists in classifying eight major skin diseases.Source: §5 · p. 12
  • The dataset was highly unbalanced and the study used no preprocessing; the authors call for more diverse, larger, balanced image sets and say image metadata may improve accuracy.Source: §5 · p. 12
  • The models could be implemented in dermoscopy systems or smartphones to assist dermatologists.Source: §5 · p. 12
Loading 1810.10348v1…