Source-linked AI summary
Knowledge Transfer for Melanoma Screening with Deep Learning
Afonso Menegola, Michel Fornaciali, Ramon Pires, Flávia Vasques Bittencourt, Sandra Avila, Eduardo Valle
TL;DR
Automated melanoma screening is challenging because melanoma is rare, lethal, rapidly progressing, and difficult to diagnose, while systematic evidence on transfer learning was missing. The paper evaluates transfer sources, fine-tuning, model depth, and lesion-label design, finding that ImageNet transfer with fine-tuning and deeper models performs best, whereas difficult lesions remain hard to diagnose.
Problem
Systematic evidence was missing on how transfer learning, its source task, fine-tuning, and model depth affect automated melanoma screening, a challenging task because melanoma is rare, lethal, rapidly progressing, and subtle to diagnose.
Method
The paper evaluates transfer from Retinopathy and ImageNet, fine-tuning, deeper architectures, and alternative choices for composing lesion classes.
Results
ImageNet transfer with fine-tuning and deeper models performed best, while difficult lesions achieved performance only little better than chance.
Takeaways & Limitations
For near-term use, referring high-confidence positive cases and difficult low-confidence cases to doctors may be more achievable than diagnosing every case with high confidence.
Abstract
from arXiv · showhide
Knowledge transfer impacts the performance of deep learning -- the state of the art for image classification tasks, including automated melanoma screening. Deep learning's greed for large amounts of training data poses a challenge for medical tasks, which we can alleviate by recycling knowledge from models trained on different tasks, in a scheme called transfer learning. Although much of the best art on automated melanoma screening employs some form of transfer learning, a systematic evaluation was missing. Here we investigate the presence of transfer, from which task the transfer is sourced, and the application of fine tuning (i.e., retraining of the deep learning model after transfer). We also test the impact of picking deeper (and more expensive) models. Our results favor deeper models, pre-trained over ImageNet, with fine-tuning, reaching an AUC of 80.7% and 84.5% for the two skin-lesion datasets evaluated.
1. INTRODUCTION
Automated melanoma screening is challenging because melanoma is rare, lethal, fast-progressing, and subtle to diagnose, while deep networks require very large datasets. The paper therefore examines transfer learning, fine-tuning, model depth, and lesion-labeling choices.
- Melanoma comprises 1% of skin-cancer cases but 75% of deaths, and prognosis worsens rapidly without early detection.
- Deep neural networks are state of the art for image classification but typically require training sets ranging from tens of thousands to millions of images.
- Existing melanoma-screening systems either train networks from scratch or transfer knowledge primarily from ImageNet, with architectures and augmentation schemes varying across studies.
- The paper avoids preprocessing such as lesion segmentation and artifact or hair removal.
- The study evaluates transfer source, fine-tuning, and model depth to clarify how knowledge transfer affects melanoma screening rather than primarily maximizing accuracy.
- It also tests whether basal cell carcinomas should be grouped with melanoma, excluded, or treated as a separate class.
2. DATA AND METHODS
The experiments train and evaluate melanoma-screening models on Atlas and ISIC while transferring from Retinopathy and ImageNet under controlled protocols. They vary transfer, fine-tuning, labeling design, and architecture, using cross-validation and AUC-based evaluation.
- Datasets: Atlas uses dermoscopic images from cases excluding acral lesions, with black image frames cropped automatically.
- Experimental design: The transfer conditions are no transfer, Retinopathy transfer, ImageNet transfer, and sequential Retinopathy-from-ImageNet transfer, each with or without fine-tuning.
- Model training: For trained or fine-tuned networks, the source output layer is replaced by a melanoma-task softmax layer and the entire network is retrained before extracting features for an SVM.
- Experimental design: The study evaluates malignant-versus-benign, melanoma-versus-benign, and three-class basal-cell-carcinoma, melanoma, and benign labeling designs.
3. RESULTS
Fine-tuning consistently improved transfer-learning performance, with ImageNet transfer outperforming Retinopathy transfer; deeper networks also performed better. Label organization and lesion difficulty materially affected screening performance.
- Fine-tuning improved classification for transfer from both Retinopathy and ImageNet, with ImageNet transfer plus fine-tuning producing the best protocol.Retinopathy transfer performed worse than ImageNet transfer, even when combined with ImageNet.
- Deeper DNN architectures performed better than shallower alternatives in the architecture comparison.
- Grouping basal cell carcinomas with melanomas or treating them as a separate class was generally easier than excluding them.The authors suggest this may reflect basal cell carcinomas being easier to diagnose than melanomas.
- Low-difficulty lesions could essentially be solved with relatively high confidence, whereas difficult lesions performed little better than chance.
4. CONCLUSIONS
The conclusions support transfer learning, fine-tuning, and deeper models, while indicating that lesion-label choices and diagnostic difficulty shape performance. The authors therefore favor referral-oriented use for difficult cases and further investigation of transfer strategies.
- Transfer learning, fine-tuning, and deeper models were supported as effective choices, but transfer from a specific related medical task posed adaptation challenges.
- Performance was sensitive to which lesions composed the positive and negative classes, possibly because melanoma and carcinoma differ in diagnostic difficulty.
- A referability framework may be more fruitful than a diagnostics framework because difficult lesions remain hard to diagnose.The proposed short-term approach refers high-confidence positive cases and low-confidence hard cases to doctors.