Application Of Artificial Intelligence In Prostate Segmentation And Cancer Invasion Identification

Prostate Cancer（PCa）is a cancer with a very high incidence in men.Early detection and treatment are key to improving patient survival.The occurrence of PCa in different region of gland is different,among which the peripheral zone accounts for70% PCa occurrence.PCa in different regions of the prostate also has different appearance on medical images.Therefore,accurate segmentation of gland region can be helpful for detection and analysis of PCa.It is more probable for PCa in boundary to break through the prostate capsule,incurring so-called extracapsular extension（ECE）.Diagnosis of ECE is very important to the surgery planning.Multi-parametric magnetic resonance imaging（mp MRI）can be used to locate PCa and assist in preoperative diagnosis of ECE.However,due to the large amount of MRI scans needs to be processed everyday,manual segmentation and diagnosis of ECE will cost radiologist much time and effort.Besides,due to the variance of image contrast brought about by mp MRI and different scan parameters,correct reading depends heavily on the experience of radiologists.This produces a large variance in the diagnosis among different radiologists.With the rapid development of computer-aided diagnosis（CAD）,artificial intelligence（AI）has been used more and more in medical imaging research.AI models can automatically extract features relevant to the problem in concern,thus can free the radiologist from heavy workload and produce more objective results.Therefore,AI models for gland sub-region segmentation and ECE classification can be established to solve the problems mentioned above.However,for both segmentation and classification,AI models need to focus on a small part of the image,such as when segmenting the small non-glandular region or focusing on the small region where ECE tends to occur for accurate diagnosis.This brings great challenges to the model building,which we improved in different ways in this study accordingly.First,for prostate sub-region segmentation,we used deep learning to segment the prostate gland to four parts,including the peripheral zone（PZ）,the transitional zone（TZ）,the distal prostatic urethra（DPU）,and the anterior fibromuscular stroma（AFS）.For PZ and TZ,satisfying results can be obtained by simple UNet.However,for DPU and AFS,the results are less than satisfactory due to their small sizes.Since both DPU and AFS may be incorrectly identified as TZ,we used an extra UNet to segment TZ from other regions besides the UNet used for fine structure segmentation,hoping to improve the segmentation of DPU and AFS by improving the TZ segmentation.Two UNet models shared some parameters.Compared with a single UNet,the proposed model not only improved the Dice of the DPU and AFS segmentation from 0.749 and0.360 to 0.760 and 0.372,respectively,but also increased the Dice of PZ and TZ to0.775 and 0.913.The performance of segmentation is close to that of the radiologists and the model can be used in valuable for analysis and diagnosis of PCa and generation of PI-RADS comformable structured reports.Second,using region of interest（ROI）of gland and PCa lesion,we proposed a method to generate an attention map to help better identify ECE of PCa.Because ECE often occurs in lesion near the edge of the gland,boundary distance was calculated between PCa and the gland to obtain a special attention map,with which radiomics and deep learning models were built.As for radiomics model,the attention map was used to extract features from background,gland and cancer,since features in normal lesion ROI could not reflect the relationship between the PCa lesion and gland.First order features of three sub-regions and the shape features of attention region were used to build the model.We compared the of the proposed model with model of prostate,PCa and the binary attention map.It was found that the features extracted from sub-regions can better reflect the characteristics of ECE and the corresponding model performed the best with the AUC（area under curve）over an internal and an external test dataset of0.792 and 0.713,respectively,which was superior to other models.We also used the aforementioned attention map in a Res Ne Xt-based network as our ECE classification model.Comparing with the whole image,the region where ECE can occur is small and may be easily ignored as the network deepens.So,self-attention module called Convolutional Block Attention Module（CBAM）was introduced into Res Ne Xt to make the model focus on the relevant region.In addition,the generated attention map was used with the CBAM,incorporating prior knowledge into the model.The model achieved AUC values in the internal and external test cohorts of 0.808 and0.713,respectively,which is better than two radiologists involved in the study,proving the value of AI model in ECE prediction.By using Grad-CAM,it was also found that the proposed model can locate the area of ECE better,which can provide visual evidences for clinicians to confirm the model findings and enhanced the interpretability of the model.