Classifying real-world macroscopic images in the primary-secondary care interface using transfer learning:...

Area of Science:

• Dermatology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) shows potential in classifying skin lesions, but generalizability across diverse populations and settings remains a challenge.
• Real-world clinical dermatology often involves macroscopic, non-standardized images, necessitating robust DL algorithms.
• Current research needs to address DL performance on varied, locally acquired datasets without extensive local labeling.

Purpose of the Study:

• To evaluate the generalizability of DL algorithms on non-dermoscopic datasets within the UK National Health Service (NHS).
• To explore methods for achieving satisfactory clinical performance using local, non-standardized data with limited local datasets.
• To quantify the impact of pretraining DL models on external, publicly available datasets.

Main Methods:

• Macroscopic diagnostic image datasets were compiled from NHS Tayside and NHS Forth Valley referrals.
• Publicly available datasets, ISIC (dermoscopic) and SD-260 (non-dermoscopic), were utilized for comparison.
• Deep learning models, including EfficientNets and SWIN transformers, were trained and fine-tuned on various data combinations, assessing performance via receiver operating characteristic curves and AUC.

Main Results:

• SWIN transformers achieved an AUC of 0.91 when pre-trained on the SD-260 dataset and fine-tuned on Forth Valley data.
• Pretraining on public macroscopic images followed by local fine-tuning improved performance for both SWIN transformers and EfficientNets across different NHS datasets.
• Pretraining on the large dermoscopic ISIC dataset did not yield additional performance benefits.

Conclusions:

• Pretraining DL models on public macroscopic images and subsequently fine-tuning on local clinical data demonstrates promising diagnostic performance.
• While promising, further enhancements are required for seamless integration into clinical workflows.
• The development of larger, domain-specific local datasets is anticipated to further boost DL algorithm performance in clinical settings.