What Is Acral Melanoma Acral melanoma (AM) occurs on the palms, soles, and nail beds - pressure-bearing areas of the hands and feet. It is the most common type of melanoma in Asian populations, but is rare in Caucasians. Because it occurs on thick skin often covered by calluses, it is frequently misidentified as a benign nevus and tends to be diagnosed late, resulting in poor outcomes.
The Clinical Challenge Dermoscopy (a non-invasive optical imaging technique) helps clinicians examine acral lesions more carefully, but its interpretation requires significant expertise. Non-experts frequently misdiagnose acral lesions because the dermoscopic patterns of acral melanoma (parallel ridge pattern, multicomponent pattern) differ markedly from patterns used to diagnose melanoma elsewhere on the body.
The AI Approach This study applied a convolutional neural network (CNN) to dermoscopy images of acral melanoma and benign acral nevi. Using transfer learning from the VGG-16 model pretrained on ImageNet, the researchers fine-tuned the network for binary classification and compared its performance to dermatologist experts and untrained physicians.
Dataset and Histopathologic Confirmation A total of 724 dermoscopy images were collected from two Korean hospitals: 350 images from 81 acral melanoma patients and 374 images from 194 benign nevus patients. All diagnoses were confirmed by histopathology. Multiple images were captured per patient for large lesions. Images included both hands/fingers (71 images) and feet/toes (653 images).
VGG-16 Architecture with Transfer Learning The team used a modified VGG-16 architecture with 16 weight layers (13 convolutional and 3 fully connected layers), replacing the final 1,000-class output layer with a 2-class output (melanoma vs. non-melanoma). Weights from layers conv1 through the first layer of conv5 were frozen using pretrained ImageNet values. The remaining layers were fine-tuned on the dermoscopy data.
Multi-Patch Sliding Window Strategy To handle variable lesion size and location, inference used a sliding window approach extracting approximately 12 image patches per test image. If any single patch was classified as melanoma, the entire image was classified as melanoma - optimizing for sensitivity. Data augmentation during training (rotation by 0, 45, 90, 135 degrees, and flips) generated 216 patches per training image, substantially expanding the small dataset.
Accuracy Comparison In Group A (training on Group B, testing on Group A), the CNN achieved 83.51% accuracy versus 81.08% for dermatologist experts and 67.84% for non-expert physicians. In Group B (training on Group A, testing on Group B), CNN accuracy was 80.23% versus 81.64% for experts and 62.71% for non-experts. CNN performance was consistently similar to expert dermatologists.
AUC and Youden's Index Both the CNN and expert dermatologists achieved AUC above 0.80 in both groups - considered 'good discrimination.' Non-experts had AUC below 0.70, falling into 'fair to poor' discrimination. Youden's index (combining sensitivity and specificity) was 0.68/0.61 for the CNN and 0.64/0.64 for experts, versus 0.39/0.25 for non-experts. The CNN thus provides expert-equivalent discrimination ability.
High Sensitivity Trade-off The CNN achieved very high sensitivity (92.57% in both groups) - slightly higher than experts in Group A and lower in Group B. However, its specificity (75-68%) was lower than non-experts (91-77%), who were more conservative in calling melanoma. The high-sensitivity design of the multi-patch strategy is appropriate for screening applications where missing a melanoma is more costly than a false positive.
Single vs. Multiple Image Input The study compared the modified VGG-16 CNN against Inception-V3 applied with both single-image and multiple-image (multi-patch) strategies. Results showed that using multiple patches improved Inception-V3 performance substantially, supporting the importance of the multi-patch strategy regardless of the underlying architecture.
Group A Results In Group A, the CNN (VGG-based) outperformed both single-image Inception-V3 (AUC 0.79) and multi-image Inception-V3 (AUC 0.81) with its AUC of 0.84. The VGG-based CNN also had the highest sensitivity (92.57%) and NPV (91.88%) in Group A.
Group B Results In Group B, the Inception-V3 models slightly outperformed the CNN in accuracy (85.19% vs. 80.23%) and AUC. The variable relative performance between groups suggests that architecture choice may be less important than dataset size and preprocessing strategy - and that larger datasets would likely improve all approaches.
Most Valuable for Non-Experts The largest performance gap between CNN and humans was against non-expert physicians. Since acral melanoma is most often encountered in primary care or general dermatology settings by clinicians who lack subspecialty dermoscopy training, a CNN-based tool could provide critical decision support - flagging suspicious acral lesions for specialist referral.
Teledermatology Application The authors note that dermoscopes connected to smartphones are increasingly used in teledermatology. A CNN-based classifier integrated into a teledermatology app could provide immediate second opinions for remote or underserved settings where specialist access is limited - potentially reducing the delayed diagnosis that characterizes acral melanoma in non-Caucasian populations.
Limitations of Acral Imaging Acral melanoma occurs on thick skin with calluses, which can obscure or distort typical dermoscopic patterns. Even expert dermatologists showed AUC around 0.82 rather than the 0.90+ seen in general melanoma studies, illustrating the inherent difficulty of this diagnostic task. Additional non-invasive imaging modalities like confocal microscopy may be needed to achieve higher accuracy.
Dataset Size Constraints With only 724 images from two Korean hospitals, the dataset is relatively small for deep learning. The 2-fold cross-validation approach is appropriate given the limitation, but larger multicenter datasets would be needed to develop a production-grade diagnostic system. Acral melanoma's low incidence even in Asian populations makes large dataset accumulation slow.
Ethnic Specificity All images were from Korean patients. Acral melanoma dermoscopic patterns may differ across ethnicities, skin phototypes, and geographic populations. Validation in other Asian cohorts (East Asian, South Asian, African descent) where acral melanoma is disproportionately prevalent would be important before generalizing the classifier.
Path Forward Combining CNN analysis with non-invasive imaging tools such as confocal microscopy or optical coherence tomography could improve accuracy by providing depth information not visible in standard dermoscopy. Additionally, multicenter collaborative databases for acral melanoma - analogous to ISIC for general melanoma - would accelerate progress in this underserved diagnostic area.