Overcoming measurement challenges in clinical practice: a deep learning-based approach to monocular surface area measurement
Published in Skin Health and Disease (Oxford Academic), this peer-reviewed study introduces a deep learning framework that accurately measures skin lesion surface areas from standard smartphone images — a critical step for objective severity scoring in conditions assessed with tools such as PASI and SCORAD.
Medela A, Sabater A, Fernández G, Mac Carthy T, Aguilar A, Herrera D, Falqués M, Martorell A. Overcoming measurement challenges in clinical practice: a deep learning-based approach to monocular surface area measurement. Skin Health and Disease. 2026. https://doi.org/10.1093/skinhd/vzag064
Figure 1: Samples compiled for evaluation
Panels (a) and (b): neck and foot samples with five ArUco markers and drawn squares of 225 mm² representing ground truth. Panels (c) and (d): corresponding samples with coloured overlays — green, yellow, and orange indicate shapes assessed with SRPE below 7%, 13%, and 20%, respectively. Error increases on surfaces lacking camera-plane alignment.
Results
Table 1: Evaluation results
| Body part | SRPE < 7% | SRPE < 13% | SRPE < 20% | MAE (mm²), mean (SD) |
|---|---|---|---|---|
| Face | 73% | 94% | 100% | 22 (18) |
| Neck | 84% | 94% | 98% | 18 (17) |
| Foot | 85% | 98% | 100% | 18 (13) |
| Total | 80% | 95% | 100% | 20 (17) |
These results show the performance of our proposed area measurement methodology when applied to different body parts. MAE, mean absolute error (lower is better); SRPE, symmetric root percentage error (higher is better).
Table 2: Performance comparison with pixel-to-area baseline
| Method | SRPE < 7% | SRPE < 13% | SRPE < 20% | MAE (mm²), mean (SD) |
|---|---|---|---|---|
| Baseline | 57% | 78% | 88% | 47 (73) |
| This study | 80% | 95% | 100% | 20 (17) |
These results show the performance of our proposed area measurement methodology in comparison with a standard methodology for area measurement. MAE, mean absolute error (lower is better); SRPE, symmetric root percentage error (higher is better).
The proposed approach reduces Mean Absolute Error by more than half (47 mm² → 20 mm²) versus conventional pixel-to-area mapping, with consistent gains across all SRPE thresholds.