Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lesion Elevation Prediction from Skin Images Improves Diagnosis

Kumar Abhishek, Ghassan Hamarneh

TL;DR

This work addresses improving skin lesion diagnosis by leveraging a latent clinical cue—lesion elevation—predicted from single 2D RGB skin images. It develops a pipeline that learns elevation from images, integrates elevation as auxiliary information into diagnosis models, and demonstrates cross-domain gains by applying elevation predictions to datasets lacking ground-truth elevation. The results show AUROC improvements of up to 6.29% for dermoscopic and 2.69% for clinical images, with significant statistics in most cases, validating both ground-truth and inferred elevations as useful cues. The findings suggest practical benefits for teledermatology and highlight future directions for dense elevation mapping and multi-dataset training to improve robustness and reduce bias.

Abstract

While deep learning-based computer-aided diagnosis for skin lesion image analysis is approaching dermatologists' performance levels, there are several works showing that incorporating additional features such as shape priors, texture, color constancy, and illumination further improves the lesion diagnosis performance. In this work, we look at another clinically useful feature, skin lesion elevation, and investigate the feasibility of predicting and leveraging skin lesion elevation labels. Specifically, we use a deep learning model to predict image-level lesion elevation labels from 2D skin lesion images. We test the elevation prediction accuracy on the derm7pt dataset, and use the elevation prediction model to estimate elevation labels for images from five other datasets: ISIC 2016, 2017, and 2018 Challenge datasets, MSK, and DermoFit. We evaluate cross-domain generalization by using these estimated elevation labels as auxiliary inputs to diagnosis models, and show that these improve the classification performance, with AUROC improvements of up to 6.29% and 2.69% for dermoscopic and clinical images, respectively. The code is publicly available at https://github.com/sfu-mial/LesionElevation.

Lesion Elevation Prediction from Skin Images Improves Diagnosis

TL;DR

This work addresses improving skin lesion diagnosis by leveraging a latent clinical cue—lesion elevation—predicted from single 2D RGB skin images. It develops a pipeline that learns elevation from images, integrates elevation as auxiliary information into diagnosis models, and demonstrates cross-domain gains by applying elevation predictions to datasets lacking ground-truth elevation. The results show AUROC improvements of up to 6.29% for dermoscopic and 2.69% for clinical images, with significant statistics in most cases, validating both ground-truth and inferred elevations as useful cues. The findings suggest practical benefits for teledermatology and highlight future directions for dense elevation mapping and multi-dataset training to improve robustness and reduce bias.

Abstract

While deep learning-based computer-aided diagnosis for skin lesion image analysis is approaching dermatologists' performance levels, there are several works showing that incorporating additional features such as shape priors, texture, color constancy, and illumination further improves the lesion diagnosis performance. In this work, we look at another clinically useful feature, skin lesion elevation, and investigate the feasibility of predicting and leveraging skin lesion elevation labels. Specifically, we use a deep learning model to predict image-level lesion elevation labels from 2D skin lesion images. We test the elevation prediction accuracy on the derm7pt dataset, and use the elevation prediction model to estimate elevation labels for images from five other datasets: ISIC 2016, 2017, and 2018 Challenge datasets, MSK, and DermoFit. We evaluate cross-domain generalization by using these estimated elevation labels as auxiliary inputs to diagnosis models, and show that these improve the classification performance, with AUROC improvements of up to 6.29% and 2.69% for dermoscopic and clinical images, respectively. The code is publicly available at https://github.com/sfu-mial/LesionElevation.
Paper Structure (6 sections, 4 equations, 4 figures, 2 tables)

This paper contains 6 sections, 4 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Results and Discussion
  4. Conclusion
  5. Acknowledgments.
  6. Disclosure of Interests.

Figures (4)

  • Figure 1: Visualizing the difference between skin lesion elevation versus depth. Invasion levels inset figure courtesy of Melanoma Institute Australia melanomadiagnosis.
  • Figure 2: derm7pt dataset: (a) sample images categorized by elevation labels and (b) distribution of elevation labels across diagnoses.
  • Figure 3: Visualizing class activation maps for skin lesion elevation label prediction for dermoscopic and clinical images, generated through GradCAM. Notice how the activation areas are focused around the lesion regions, indicating that the prediction model $g$ does not learn to rely on spurious features or "shortcuts".
  • Figure SM1: Dense depth maps for dermoscopic and clinical images from derm7pt generated using MiDaS ranftl2022towards, an off-the-shelf monocular depth estimation model1. Notice how the scene anisotropy of the images that this model has been trained on shows up in the generated depth maps - the lower regions of natural images are almost always closer to the camera plane, and this is reflected in these predicted depth maps. However, that is generally not the case with skin lesion images, which are typically acquired by having the camera plane parallel to the skin surface. This, along with the difference in scale between natural images' depths (typically in meters) and skin lesion elevations (typically in millimeters), considerably limits the utility of these depth map estimates.