1955P - Distinguishing lipomas from atypical lipomatous tumors/well differentiated liposarcomas in histology slides using deep learning

Background

Liposarcomas are fat tissue tumors with an incidence rate of 1 case per 100,000 population per year. Atypical Lipomatous Tumors / Well Differentiated Liposarcomas (ALT/WDLPS) are malignant tumors accounting for around 40-45% of liposarcomas and are particularly challenging to diagnose as they can be easily confused with lipomas, another tumor of mesenchymal origin which is benign. The current diagnosis protocol consists in producing a histology slide from a tumor sample (either obtained by resection or biopsy) so that pathologists can determine the type of tumor. If the visual inspection remains uncertain, or the clinical suspicion of ALT/WDLPS is high, a FISH test is ordered in order to analyze the status of MDM2 gene.

Methods

This is a study on the ability of modern deep learning based AI technologies to automatically classify a whole slide image (WSI) containing fat tissue tumor into lipoma or ALT/WDLPS. Two cohorts were provided by Institut Bergonié (IB), Bordeaux, France and Centre Léon Bérard (CLB), Lyon, France. The IB cohort contains 1103 slides (536 Lipomas / 567 ALT/WDLPS) while the CLB cohort contains 956 slides (668 Lipomas / 288 WDLPS). We propose to use a deep learning pipeline relying on (i) a process mapping a WSI into 4000 (112x112 micrometers) small areas randomly sampled from regions containing matter, (ii) a transformer network that maps image areas into a 768 dimensional vectors and (iii) a multiple instance learning (MIL) head for the computation of a probability of WDLPS at the slide level.

Results

We trained our deep learning model using a 4-fold cross-validation (CV) with 5 repeats. Evaluation is reported in terms of accuracy (ACC), i.e., correct prediction rate. In the IB cohort, the model achieves 84.7% of ACC (std 2.5%). In the CLB cohort, it achieves 90.5% of ACC (std 2.1%).

Conclusions

The trained model exhibits excellent performances on (held-out) test samples in both cohorts. Many improvements can be made to the model by either adding clinical data in the pipeline or learning a multi-cohort model through federated learning.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Owkin France.

Funding

Owkin France.

Disclosure

A. Camara, J. Klein, T. Marchand, C. Maussion, V. Rousseau, T. Bonnotte: Financial Interests, Institutional, Member: Owkin France. All other authors have declared no conflicts of interest.