Abstract PO-025: Differentiation of benign from clinically significant prostate cancer tissues using convolution neural networks on raw micro-ultrasound data

Abstract Purpose: Micro-ultrasound is a promising new technique offering high-resolution images to identify prostate cancer. However, clinical interpretation of the images remains challenging. The underlying raw radio-frequency ultrasound is a unique source of tissue acoustic properties, potentially ideal to identify tissues-of-interest in images for targeted biopsies. Here, we develop a convolutional neural network (3D CNN) to classify tissue as benign vs clinically significant prostate cancer (csPCa) using the raw RF data. Materials and Methods: A total of 837 male patients (median age 63, IQR 57-68) suspected of prostate cancer and undergoing a prostate biopsy were recruited. All data was obtained through an IRB approved at the local data acquisition site (5 collaborating sites); all patients were consented before 29MHz micro-ultrasound (ExactVu, Exact Imaging, Markham, Canada) data acquisition during targeted biopsy. Patients across all sites were scanned using consistent acquisition presets, and all data was saved in raw IQ format. Up to 12 IQ images were obtained for each patient, representing targeted biopsy locations. Histopathological analysis of each biopsy sample served as the clinical standard of benign vs clinically significant cancer (Gleason score (GS) ≥7). Samples containing clinically insignificant cancer (GS6) were not analyzed. Data was processed to maintain both spatial and frequency information for images. A shallow 3D CNN model was trained on patient images (train-test split based on patients); a similar network was also trained with the addition of Prostate Specific Antigen (PSA) level for each patient as a feature. Training used 5220 patient images (4520 benign vs. 700 GS7+;80% of the data); testing was carried out on a set-aside data set of 1310 patient images (1131 benign vs. 179 GS7+ images; 20% of data). The area area under the receiver operator curve (ROC-AUC) was the main evaluation metric. We also explored whether our models could help classify data on a patient level of benign vs. csPSa patients. Results: Preliminary results suggest that our model yields a ROC-AUC of 82% on an image-level to differentiate between benign and csPSa-confirmed images. Including the PSA into our model during training resulted in ROC-AUCs of up to 85%. On a patient level, up to 81% ROC-AUC was achievable in differentiating patients with benign vs. csPSa. Including the PSA on a patient-level resulted in an 87% ROC-AUC. Conclusion: CNNs can help capture unique tissue acoustic properties in micro-ultrasound images to help identify areas suspicious for prostate cancer that warrant targeted biopsy. This offers the potential to improve cancer diagnosis while reducing the number of biopsy cores required. Citation Format: Ahmed El Kaffas, Adi Lightstone, Raul Garcia, Brian Woodlinger, Miriam Ibrahim, Richard Fan, Geoffrey Sonn. Differentiation of benign from clinically significant prostate cancer tissues using convolution neural networks on raw micro-ultrasound data [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-025.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.