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

Congress:

ECR 2022

Poster Number:

C-22165

Type:

Scientific Exhibit

Keywords:

Artificial Intelligence, Genital / Reproductive system male, Oncology, MR, MR-Diffusion/Perfusion, Biopsy, Computer Applications-Detection, diagnosis, Cancer

Authors:

N. Debs, A. Routier, B. Lorenzi, L. Wood, F. Nicolas, M.-M. Rohé

DOI:

10.26044/ecr2022/C-22165

DOI-Link:

https://dx.doi.org/10.26044/ecr2022/C-22165

Conclusion

Our experiments show the interest of combining radiological annotation (PIRADS) with histopathological annotation (Gleason score), in order to improve computer aided detection and diagnosis systems of prostate cancer
Detection results from our model trained on PIRADS and GS annotations is competitive with respect to the best performances found in state of the art:
- Saha et al.: 83% of CSC lesions detected at 0.5 FP/volume, 93% at 1FP/volume. Evaluation done on personal evaluation data [9]
- Bosma et al.: 84% of CSC lesions detected at 1FP/volume. Evaluation done on personal data [10]
- Yu et al: 85% of CSC lesion detected at 0.62 FP/volume, 89% at 0.94 FP/volume. Evaluation done on a subset of ProstateX dataset [11]
Classification performance reflects the difficulty of the classification task. We seek to discriminate CSC lesions from benign lesions. Only among benign lesions, we include in the training database both lesions annotated GS<7 as well as lesions annotated PIRADS 3. The ambiguity and the intra- and inter-expert variability for PIRADS 3 lesions [12,13] may explain the relatively low AUC observed

Fig 1: Two possible annotation types (PIRADS score and Gleason score)

Fig 2: Annotation tool for PIRADS scoring

Fig 3: Score distribution for lesions in the training/validation dataset with both...

Fig 4: Overview of the deep learning network

Fig 5: Overview of datasets used for our experiments

Fig 6: FROC analysis for detection sensitivity on CS lesions (GS>6), depending on the...

Fig 7: Example of a patient predicted by both models (PIRADS+Gleason model and Gleason...

Fig 8: FROC analysis for detection sensitivity on CS lesions (GS>6), depending on the...