Enc‐Unet: A novel method for Glioma segmentation. Issue 2 (5th November 2022)

• Record Type: Journal Article
• Authors: Hussain, Syed Sajid; Sachdeva, Jainy; Ahuja, Chirag Kamal; Singh, Abhiav
• Is Part Of: International journal of imaging systems and technology. Volume 33:Issue 2(2023)
• Page Start: 465
• Page End: 482
• Publication Date: 2022-11-05
• Subjects: 3D‐Unet -- autoencoders -- deep learning -- glioma segmentation -- MRI; Imaging systems -- Periodicals; Image processing -- Periodicals; 621.367
• DOI: 10.1002/ima.22822
• Languages: English
• ISSNs: 0899-9457

Abstract

Abstract: The diagnosis' treatment planning, follow‐up and prognostication of Gliomas is significantly enhanced on Magnetic Resonance Imaging. In the present research, deep learning‐based variant of convolutional neural network methodology is proposed for glioma segmentation where pretrained autoencoder acts as backbone to the 3D‐Unet which performs the segmentation task as well as image restoration. Further, Unet accepts input as the combination of three non‐native MR images (T2, T1CE, and FLAIR) to extract maximum and superior features for segmenting tumor regions. Further, weighted dice loss employed, focusses on segregating tumor region into three regions of interest namely whole tumor with oedema (WT), enhancing tumor (ET), and tumor core (TC). The optimizer preferred in the proposed methodology is Adam and the learning rate is initially set to 1 e − 4, progressively reduced by a cosine decay after 50 epochs. The learning parameters are reduced to a larger extent (up to 9.8 M as compared to 27 M). The experimental results show that the proposed model achieved Dice similarity coefficients: 0.77, 0.92, and 0.84; sensitivity: 0.90, 0.95, and 0.89; specificity: 0.97, 0.99, and 0.99; Hausdorff95: 5.74, 4.89, and 6.00, in the three regions including ET, WT, TC. This proposed Glioma segmentation method is efficient for segregation of tumors. Significance: The 3D tumor‐Glioma segmentation is a voxel‐based problem where the tumor is segmented into whole tumor with oedema, enhancing tumor, and tumor core on magnetic resonance images. In the present study, pretrained autoencoder is developed which acts as an anchor for 3D‐Unet for segmenting the tumor boundaries and restoring of an image. Further, during pre‐training, the loss function namely "Mean Squared Error" loss function procured a restored image. The false positive as well as false negative samples during segmentation are obtained due to consideration of improved weighted loss dice function. The learning parameters reduced to a to 9.8 M as compared to 27 M from state of art methods. This resulted in better output visually as well as in terms of other statistical parameters such as Dice similarity coefficient, specificity, sensitivity, and Hausdorff95.