Deep learning-based dot and globule segmentation with pixel and blob-based metrics for evaluation

Abstract

Deep learning (DL) applied to whole dermoscopic images has shown unprecedented accuracy in differentiating images of melanoma from benign lesions. We hypothesize that accuracy in whole-image deep learning suffers because whole lesion analysis lacks an evaluation of dermoscopic structures. DL also suffers a “black box” characterization because it offers only probabilities to the physician and no visible structures. We propose the detection of structures called dots and globules as a means to improve precision in melanoma detection. We compare two encoder-decoder architectures to detect dots and globules: UNET vs. UNET++. For each of these architectures, we compare three pipelines: with test-time augmentation (TTA), without TTA, and without TTA but with checkpoint ensembles. We use an SE-RESNEXT encoder and a symmetric decoder. The pixel-based F1-scores for globule and dot detection based on UNET++ and UNET techniques with checkpoint ensembles were found to be 0.632 and 0.628, respectively. The blob-based UNET++ and UNET F1-scores (50 percent inter-section) were 0.696 and 0.685, respectively. This agreement score is over twice the statistical correlation score measured among specialists. We propose UNET++ globule and dot detection as a technique that offers two potential advantages: increased diagnostic accuracy and visible structure detection to better explain DL results and mitigate deep learning's black-box problem. We present a public globule and dot database to aid progress in automatic detection of these structures.

Keywords

• Machine learning
• Deep learning
• Data processing
• Melanoma
• Globules
• Feature segmentation