Contribution of PIRADS annotation for the detection of prostate significant cancer using deep learning

International audience Introduction: Gleason score (GS) is the most powerful prognostic factor for prostate cancer. Although this histopathological score is a reliable measure of cancer aggressiveness, its obtention is quite invasive for patients via biopsy. Therefore, it is not routinely determined when cancer is suspected. The PIRADS radiological score, on the other hand, is a pre-biopsy score based on multi-parametric MRI that is much less invasive and relatively reliable in predicting aggressiveness of prostate cancer. Today, majority of studies proposing automatic systems for prostate cancer detection and diagnosis are only based on GS annotated data, aiming to focus on reliable annotation. We show through this study the interest of adding PIRADS annotated data in addition to GS annotated data, to increase the detection and classification performances of clinically significant lesions. Material and Method: We had a bi-parametric, multi-center MRI database from various continents. We performed two experiments, for which two distinct deep learning models were trained: a first model trained only on data for which GS annotation was available (i.e. 442 training patients), and a second model trained both on the data with GS and PIRADS annotations (i.e. 2473 training patients). For this second model, we hypothesize that PIRADS score 3 is equivalent to GS=<6 and that PIRADS score 4-5 is equivalent to GS>6. Both models were evaluated on the same samples for which GS annotation was available, namely 140 validation data and 98 test data from the ProstateX2 challenge. Detection performance was measured using the FROC curve, and classification performance (GS>6 vs GS<=6) using the AUC.Results:On the validation set, the model trained on PIRADS and GS data had a percentage of significant lesions (GS>6) detected of 0.74 at 0.5 FP/volume and 0.79 at 1FP/volume, compared with 0.68 at 0.5 FP/volume and 0.72 at 1FP/volume for the model trained on GS data only. The calculated AUC was 0.723 for the model trained on PIRADS and GS data, versus 0.699 for the model trained on GS data only. On the test set, the model trained on PIRADS and GS data had a percentage of significant lesions (GS>6) detected of 0.63 at 0.5 FP/volume and 0.68 at 1FP/volume, compared with 0.60 at 0.5 FP/volume and 0.61 at 1FP/volume for the model trained on GS data only. The calculated AUC was 0.612 for the model trained on PIRADS and GS data, versus 0.558 for the model trained on GS data only. Conclusion:We show the interest of combining radiological annotation with histopathological annotation, in order to improve computer aided detection and diagnosis systems of prostate cancer.