Study Goal Researchers at the National Cancer Center in Japan developed a deep learning classification system specifically designed to distinguish malignant from benign pigmented skin lesions using standard clinical photographs - not specialized dermoscopy images.
Clinical Motivation Melanoma accounts for the majority of skin cancer deaths worldwide, and in Japan specifically, patients often present at advanced stages because melanoma is less common and awareness is low. The average delay from first noticing a lesion to visiting a doctor in Japan is 69.5 months - far longer than the 2-9.8 months reported in other countries.
Key Innovation Unlike prior AI skin cancer studies that used dermoscopic images, this system uses everyday clinical camera photographs - the kind of image any member of the public or general practitioner might take. This makes the technology more accessible for broad deployment.
Main Finding The AI system (called FRCNN) achieved 86.2% accuracy on six-class classification, significantly outperforming both board-certified dermatologists (79.5%) and trainees (75.1%). In benign-versus-malignant classification, FRCNN reached 91.5% accuracy versus 86.6% for specialists.
Image Collection From a pool of over 120,000 clinical photographs taken at the National Cancer Center Hospital between 2001 and 2017, researchers extracted 5,846 images of brown-to-black pigmented lesions from 3,551 patients. All images were taken with high-quality digital SLR cameras using macro lenses and ring flashes.
Disease Categories The dataset covered six classes: malignant melanoma (1,611 images), basal cell carcinoma (401 images), nevus (2,837 images), seborrheic keratosis (746 images), senile lentigo (79 images), and hematoma/hemangioma (172 images). All malignant tumors were histopathologically confirmed by biopsy.
Train/Test Split The researchers randomly selected 666 patients (one image each) for the test dataset and used the remaining 4,732 images for training. To limit burden on dermatologists, each was given a subset of 200 images per test session, with 10 test sets created so each image appeared at least three times across all evaluations.
Faster R-CNN Architecture The Faster Region-based Convolutional Neural Network (FRCNN) was chosen for its combination of speed and accuracy. It uses a VGG-16 backbone and integrates region proposal and classification into a single unified network. Training included data augmentation techniques such as rotation, flipping, random cropping, and zoom.
Optimization Settings The FRCNN was trained using momentum stochastic gradient descent with a learning rate of 0.001, momentum of 0.9, and weight decay of 0.0005. Training ran for 100 epochs, with the learning rate decreased by a factor of 10 at epochs 40 and 80.
Handling Class Imbalance Because basal cell carcinoma, senile lentigo, and hematoma/hemangioma had fewer images than other categories, these classes were oversampled twice during training. This helped prevent the model from becoming biased toward more common classes.
Test-Time Augmentation During inference, each input image was transformed 30 different ways - combining horizontal flips (2 patterns), 72-degree rotations (5 patterns), and three zoom levels. Predictions from all 30 versions were evaluated and the highest-confidence prediction was selected as the final result.
Implementation The model was implemented using Chainer, ChainerCV, and CuPy libraries. This Japanese deep learning ecosystem was well-suited for the GPU-accelerated training required for the large image dataset.
Six-Class Classification FRCNN achieved 86.2% accuracy across all six skin lesion types, significantly better than board-certified dermatologists (79.5%, p=0.0081) and trainees (75.1%, p less than 0.00001). In 9 of 10 test sets, the AI outperformed all human examiners.
Two-Class Classification When classifying lesions as simply benign or malignant, FRCNN reached 91.5% accuracy compared to 86.6% for specialists and 85.3% for trainees. These differences were statistically significant (p=0.0083 and p less than 0.001, respectively).
False Positive Rates A critical finding was that dermatologists had much higher false positive rates (13.4% for specialists, 14.1% for trainees) compared to FRCNN (5.5%). When uncertain, human doctors tended to classify lesions as malignant, leading to more unnecessary biopsies or patient anxiety.
Consistency The AI showed much lower variability across test sessions (standard deviation of 1.69%) compared to dermatologists (up to 9.79% for specialists). This robustness is a practical advantage for real-world deployment where consistent performance matters.
Why AI Outperforms Human dermatologists, when uncertain, tend to classify lesions as malignant - a conservative strategy to avoid missed cancers, but one that inflates false positives. AI does not suffer from fatigue, loss of concentration, or variability in clinical experience that affects human examiners.
Use Case for General Public The authors envision this system deployed on smartphones or wearable devices to help non-medical users identify potentially cancerous skin lesions. The goal would be encouraging timely doctor visits rather than serving as a definitive diagnosis tool.
Mobile Application Challenges Existing smartphone apps for skin cancer detection have low accuracy partly because image quality varies dramatically between devices and shooting conditions. The authors acknowledge these as key barriers to overcome for widespread deployment.
Japan-Specific Problem In Japan, there is no official skin cancer screening program, and melanoma is relatively rare, leading to low awareness. The long delay before patients seek care contributes to late-stage presentations and poor prognosis - exactly the gap this technology aims to address.
Benign Diagnosis Uncertainty While all malignant tumors were confirmed histopathologically, benign tumors were diagnosed clinically using dermoscopy - without biopsy in most cases. This introduces some uncertainty in the ground-truth labels used for training and evaluation.
Single-Institution Data The training dataset came entirely from one hospital, using consistent camera equipment and imaging protocols. The system's generalizability to images taken with different cameras, lighting, or angles - as would be common in community settings - has not been validated.
Algorithm Limitations FRCNN has known difficulties with low-resolution images due to its limited capacity for identifying local texture. As deep learning technology advances rapidly, the authors plan to evaluate newer architectures to improve performance.
Next Steps The authors plan to build a large-scale social implementation dataset capturing diverse image conditions, patient demographics, and lesion characteristics to improve the system over time and enable deployment for general public use.