Recall-Oriented Adaptive Loss with Knowledge Distillation for Efficient Melanoma Detection Under Class Imbalance

Melanoma is a highly aggressive skin cancer for which early and accurate detection remains essential. However, automated analysis of dermoscopic images is challenged by severe class imbalance, which can bias deep learning models toward the non-melanoma class and reduce sensitivity. This study presents a recall-oriented knowledge distillation framework that combines the Recall-Oriented Adaptive Loss (ROAL) with teacher–student learning. ROAL increases the positive-class weight whenever melanoma recall on a development subset falls below a predefined target, while a ResNet50 teacher transfers softened predictions to a MobileNetV2 student with approximately 2.2 million parameters, compared with 23.5 million for the teacher. Using a lesion-level split, three random seeds, and a fixed protocol, eleven teacher, student-only, and distillation configurations were evaluated on HAM10000 and further assessed on the more severely imbalanced ISIC 2020. At the common predefined operating threshold of 0.5, distilling a ROAL-trained teacher into a downsampled student achieved the highest melanoma recall on both datasets: 0.924 on HAM10000, with an average of seven false negatives out of 92 melanoma cases, and 0.914 on ISIC 2020, with 9.7 out of 112. In both cases, the distilled student exceeded the recall of its own teacher while using ten times fewer parameters, a gain that reflects the combination of distillation with balanced-batch training rather than distillation alone. On the second dataset, ROAL showed a larger performance gap over fixed class weighting, indicating that the proposed adaptive strategy remained effective under a substantially different and more imbalanced data distribution. Overall, the framework achieves high melanoma sensitivity at a fraction of the model size, supporting its potential deployment in resource-constrained screening settings.

Authors

Institutions

Publication Details

Journal
Machine Learning and Knowledge Extraction
Published
2026-11-17
DOI
https://doi.org/10.3390/make8110384
Primary Topic
Cutaneous Melanoma Detection and Management
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Recall-Oriented Adaptive Loss with Knowledge Distillation for Efficient Melanoma Detection Under Class Imbalance

Samah Elkah, My Abdelouahed Sabri, Abdellah Aarab
4 links
Machine Learning and Knowledge Extraction
Cutaneous Melanoma Detection and Management
article

Recall-Oriented Adaptive Loss with Knowledge Distillation for Efficient Melanoma Detection Under Class Imbalance

Samah Elkah, My Abdelouahed Sabri, Abdellah Aarab
article en

Abstract

Melanoma is a highly aggressive skin cancer for which early and accurate detection remains essential. However, automated analysis of dermoscopic images is challenged by severe class imbalance, which can bias deep learning models toward the non-melanoma class and reduce sensitivity. This study presents a recall-oriented knowledge distillation framework that combines the Recall-Oriented Adaptive Loss (ROAL) with teacher–student learning. ROAL increases the positive-class weight whenever melanoma recall on a development subset falls below a predefined target, while a ResNet50 teacher transfers softened predictions to a MobileNetV2 student with approximately 2.2 million parameters, compared with 23.5 million for the teacher. Using a lesion-level split, three random seeds, and a fixed protocol, eleven teacher, student-only, and distillation configurations were evaluated on HAM10000 and further assessed on the more severely imbalanced ISIC 2020. At the common predefined operating threshold of 0.5, distilling a ROAL-trained teacher into a downsampled student achieved the highest melanoma recall on both datasets: 0.924 on HAM10000, with an average of seven false negatives out of 92 melanoma cases, and 0.914 on ISIC 2020, with 9.7 out of 112. In both cases, the distilled student exceeded the recall of its own teacher while using ten times fewer parameters, a gain that reflects the combination of distillation with balanced-batch training rather than distillation alone. On the second dataset, ROAL showed a larger performance gap over fixed class weighting, indicating that the proposed adaptive strategy remained effective under a substantially different and more imbalanced data distribution. Overall, the framework achieves high melanoma sensitivity at a fraction of the model size, supporting its potential deployment in resource-constrained screening settings.

Machine Learning and Knowledge ExtractionVol. 8(11)
Sidi Mohamed Ben Abdellah University (MA)
Good health and well-being
Cutaneous Melanoma Detection and Management
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon