Improving Automatic Melanoma Diagnosis Using Deep Learning-Based Segmentation of Irregular Networks

SIMPLE SUMMARY: Supervised deep learning techniques can now automatically process whole dermoscopic images and obtain a diagnostic accuracy for melanoma that exceeds that of specialists. These automatic diagnosis systems are now appearing in clinics. However, the computational techniques used cannot be easily interpreted by the experts using the systems, and they still fail to detect a minority of melanomas. We describe an approach that detects critical irregularities in pigment patterns, a clinical feature, and uses this knowledge to improve deep learning diagnostic accuracy. In this research, we trained a deep learning network to identify visible patterns associated with melanoma. We combine these patterns with a supervised whole-image deep learning method to improve diagnostic accuracy and provide a publicly available dataset with the clinical structures annotated. ABSTRACT: Deep learning has achieved significant success in malignant melanoma diagnosis. These diagnostic models are undergoing a transition into clinical use. However, with melanoma diagnostic accuracy in the range of ninety percent, a significant minority of melanomas are missed by deep learning. Many of the melanomas missed have irregular pigment networks visible using dermoscopy. This research presents an annotated irregular network database and develops a classification pipeline that fuses deep learning image-level results with conventional hand-crafted features from irregular pigment networks. We identified and annotated 487 unique dermoscopic melanoma lesions from images in the ISIC 2019 dermoscopic dataset to create a ground-truth irregular pigment network dataset. We trained multiple transfer learned segmentation models to detect irregular networks in this training set. A separate, mutually exclusive subset of the International Skin Imaging Collaboration (ISIC) 2019 dataset with 500 melanomas and 500 benign lesions was used for training and testing deep learning models for the binary classification of melanoma versus benign. The best segmentation model, U-Net++, generated irregular network masks on the 1000-image dataset. Other classical color, texture, and shape features were calculated for the irregular network areas. We achieved an increase in the recall of melanoma versus benign of 11% and in accuracy of 2% over DL-only models using conventional classifiers in a sequential pipeline based on the cascade generalization framework, with the highest increase in recall accompanying the use of the random forest algorithm. The proposed approach facilitates leveraging the strengths of both deep learning and conventional image processing techniques to improve the accuracy of melanoma diagnosis. Further research combining deep learning with conventional image processing on automatically detected dermoscopic features is warranted.