Automatic Skin Lesion Analysis using Large-scale Dermoscopy Images and Deep Residual Networks (1703.04197v2)

Abstract: Malignant melanoma has one of the most rapidly increasing incidences in the world and has a considerable mortality rate. Early diagnosis is particularly important since melanoma can be cured with prompt excision. Dermoscopy images play an important role in the non-invasive early detection of melanoma [1]. However, melanoma detection using human vision alone can be subjective, inaccurate and poorly reproducible even among experienced dermatologists. This is attributed to the challenges in interpreting images with diverse characteristics including lesions of varying sizes and shapes, lesions that may have fuzzy boundaries, different skin colors and the presence of hair [2]. Therefore, the automatic analysis of dermoscopy images is a valuable aid for clinical decision making and for image-based diagnosis to identify diseases such as melanoma [1-4]. Deep residual networks (ResNets) has achieved state-of-the-art results in image classification and detection related problems [5-8]. In this ISIC 2017 skin lesion analysis challenge [9], we propose to exploit the deep ResNets for robust visual features learning and representations.

Analysis of Automatic Skin Lesion Analysis using Deep Residual Networks

The paper "Automatic Skin Lesion Analysis using Large-scale Dermoscopy Images and Deep Residual Networks" presents a comprehensive approach to the detection and classification of malignant melanoma using state-of-the-art deep learning technologies. Given the high incidence and mortality rates associated with melanoma, early diagnosis facilitated by dermoscopy images is of critical importance. However, the subjective nature of visual assessment by clinicians necessitates automation to ensure accuracy and reproducibility. This research explores the use of Deep Residual Networks (ResNets) for the automatic analysis of dermoscopy images, aiming at both lesion segmentation and classification.

Methodology

The authors utilize ResNets, known for their efficacy in image classification tasks, overcoming the limitations associated with the depth of many traditional networks like VGGNet. The architecture leverages residual blocks, which introduce shortcut connections, thus enabling effective training of very deep networks. This structural innovation allows ResNets to behave as ensembles of shallower networks, improving feature learning and representation for robust skin lesion analysis.

Lesion Segmentation

The segmentation process employs a Fully Convolutional Networks (FCN) architecture to upsample features from ResNet, producing score masks for lesion boundaries. A significant dataset, composed of approximately 9,800 training images, supports this task, with images obtained from the ISIC archive and tuned via innovative data augmentation techniques. The multi-scale integration approach employed during testing proved crucial for performance enhancement.

Lesion Classification

The classification task is divided into melanoma, seborrheic keratosis, and nevus categories. Three classification strategies are pursued: multi-class classification, binary classification, and an ensemble approach combining the former two. A separate dataset of approximately 3,600 images forms the basis for training, again leveraging pre-trained ResNets fine-tuned for this specific task.

Experimental Results

The proposed methods were evaluated using a validation set provided by the ISIC 2017 challenge, which included 150 dermoscopic images with undisclosed ground truths to the participants.

• Segmentation: The ResNet-Segmentation method, especially in combination with multi-scale integration (MResNet-Seg), achieved superior performance, evidenced by a 79.4% Jaccard index with additional training images. This performance positioned the approach first on the validation set leaderboard.
• Classification: The ensemble classification approach demonstrated the highest efficacy, achieving an average AUC of 91.5% across melanoma and seborrheic keratosis cases, outperforming both single multi-class and binary classification strategies.

Implications and Future Research

This paper underscores the potential of deploying deep learning models, particularly ResNets, in medical image analysis for dermatology. With demonstrated improvements in lesion segmentation and classification, these methods could significantly aid clinical decision-making by providing reliable diagnostic outputs.

The results suggest several areas for future research, including optimizing network architectures for better computational efficiency, exploring ensemble methods further, and testing on more varied datasets to ensure broad generalizability. Additionally, integrating other modalities such as histopathology images or patient metadata into the analytical framework could augment the predictive capabilities of these models. As AI continues to evolve, its integration into dermatological diagnostics promises to enhance patient outcomes through precision medicine.