FusionM4Net Uses a Two-Stage Approach to Boost Diagnostic Accuracy in Skin Lesion Classification
Designed around a two-prong fusion approach — combining image data first then metadata second — FusionM4Net has 78.5 percent accuracy.
A team led by PD Dr. Tobias Lasser from the Technical University of Munich (TUM) has developed a new machine learning system for accurate automated classification of skin lesions — and the code has been released under an unspecified open source license for all to investigate.
"Future routine clinical use of algorithms with high diagnostic accuracy might help ensure that rare diseases are also detected by less experienced physicians," Lasser claims of the work of his team and the impact of future projects yet to come, "and it might mitigate decisions affected by stress or fatigue."
The key to Lasser's statement: "High diagnostic accuracy." Without that, there's the risk of false diagnosis — either claiming to have found an issue which on closer inspection is not there or, worse, missing an issue which could be resolved with early enough intervention.
That's where Lasser's team's machine learning system comes in. Dubbed FusionM4Net, the platform is designed to use multi-modal multi-stage data fusion: It integrates data from clinical images, microscope images of suspicious skin lesions, and patient metadata; it differentiates between five different categories of skin lesions; and it works across two key stages, merging the available image data first then adding the metadata in a second stage — a key differentiation, its creators claim, from rival single-stage approaches.
The result is a high-accuracy system: Following training on a publicly-available dataset, the average diagnostic accuracy was measured at 78.5 percent — outperforming all other algorithms against which it was tested, though falling short of guaranteed accuracy.
The team is now working on readying the algorithm for clinical deployment, teaming up with the Department of Dermatology and Allergology at the University Hospital of LMU Munich to integrate a range of standardized but non-public datasets.
The team's work has been published in the journal Medical Image Analysis under closed-access terms; the raw Python source code and Jupyter notebook for FusionM4Net has been published to GitHub under an unspecified open source license.
Main article image courtesy of LMU Klinikum München; Lasser left.