Prostate  cancer detection using Artificial Intelligence in MRI


Year:

Session type:

Theme:

Reddy Reddy, Dr.sanjai Addla, Dr.raghu Nath

Abstract

Background

Since the precise diagnosis of tumors plays a crucial role in staging and treatment plans, we need an efficient and accurate software to help cancer physicians to perform deep diagnosis of prostate cancer further can help in right treatment for the patient. To assist such complex task we are developing deep learning software for automated segmentation of prostate gland and segment tumor to score Gleason score in 3D MRI volumes

Method

 In this work we propose an approach to 3D image segmentation based on a volumetric, fully convolutional, neural network. Our CNN is trained end-to-end on MRI volumes depicting prostate and learns to predict segmentation for the whole volume at once. 1000 training sets and 200 validation sets were prepared. Patches of size 64 X 64 X 12 pixels were extracted around the finding. MRI dataset of about 250 patients was prepared and the CNNs was designed using Keras neural-network. In order to train our model using the 3-dimensional magnetic resonance images, we must use a volumetric convolution, taking into account every layer –Axial, Coronal and Sagittal dimensions.

Results

Prostate tumor segmentation and Gleason scoring

Patients with prostate cancer with a Gleason score of 7 or higher were considered clinically significant cancers. The CNN was trained on MRI images of 150 patients, using the segmentation performed by expert reader. For each patient, patches (size M × M = 21 × 21 voxels) were created by combining T2-weighted, and DWI images, for both tumor and non-tumor areas. The discovery data comprised an independent discovery set (totaling 7000 patches) and a test set (2000 patches). The accuracy was 0.85­.

Conclusion

Conclusions: We presented an approach based on a volumetric convolutional neural network that performs segmentation of MRI prostate volumes in a fast and accurate manner and implements deep learning-based tools for automatic classification of the cancerous findings in multi-parametric MRIs.

Impact statement

Value for Radiologists/Oncologists- Save Time, evidences for confidence, performance, second opinion, the bottom of the pyramid.