Ensemble U‐net‐based method for fully automated detection and segmentation of renal masses on computed tomography images

Abstract

Purpose: Detection and accurate localization of renal masses (RM) are important steps toward future potential classification of benign vs malignant RM. A fully automated algorithm for detection and localization of RM may eliminate the observer variability in the clinical workflow. Method: In this paper, we describe a fully automated methodology for accurate detection and segmentation of RM from contrast-enhanced computed tomography (CECT) images. We first determine the boundaries of the kidneys on the CECT images utilizing a convolutional neural network-based method to be used as a region of interest to search for RM. We then employ a homogenous U-Netbased ensemble learning model to identify and delineate RM. We used an institutional dataset comprised of CECT images in 315 patients to train and evaluate the proposed method. We compared results of our method to those of three-dimensional (3D) U-Net for RM localization and further evaluated our algorithm using the kidney tumor segmentation (KiTS19) challenge dataset. Results: The developed algorithm reported a Dice similarity coefficient (DSC) of 95.79% AE 5.16% and 96.25 AE 3.37 (mean AE standard deviation) for segmentation accuracy of kidney boundary from 125 and 60 test images from institutional and KiTS19 datasets, respectively. Using our method, RM were detected in 125 and 52 test cases, which corresponds to 100% and 86.67% sensitivity at patient level in institutional and KiTS19 test images. Our ensemble method for RM localization yielded a mean DSC of 88.65% AE 7.31% and 87.91% AE 6.82% on the institutional and KiTS19 test datasets, respectively. The mean DSC for RM delineation from CECT institutional test images using 3D U-Net was 85.95% AE 1.46%. Conclusion: We describe a method for automated localization of RM using CECT images. Our results are important in terms of clinical perspective as fully automated detection of RM is a fundamental step for further diagnosis of cystic vs solid RM and eventually benign vs malignant solid RM, that has not been reported previously.

Unified Focal loss: Generalising Dice and cross entropy-based losses to handle class imbalanced medical image segmentation

“…This framework was optimised using the Dice and cross entropy loss. Recently, an ensemble-based method obtained comparable results to nnU-Net, and involved initial independent processing of kidney organ and kidney tumour segmentation by 2D U-Nets trained using the Dice loss, followed by suppression of false positive predictions of the kidney tumour segmentation using the network trained for kidney organ segmentation ( Fatemeh et al, 2020 ). When the dataset size is small, results from an active learning-based method using CNN-corrected labelling, also trained using the Dice loss, showed a higher segmentation accuracy over nnU-Net ( Kim et al, 2020 ).…”

Deep Learning Algorithm for Fully Automated Detection of Small (≤4 cm) Renal Cell Carcinoma in Contrast-Enhanced Computed Tomography Using a Multicenter Database

“…In this study, we developed an algorithm for the fully automatic detection of small RCCs in CECT images based on deep learning. The algorithm yielded comparable sensitivity and specificity to those of previous studies 15,16 . To the best of our knowledge, our study is the first to develop an algorithm for the detection of small RCCs from CECT images and validate it using a large number of scans from a multicenter database.…”

The Application and Development of Deep Learning in Radiotherapy: A Systematic Review

“…Compared with the gold standard, using 3D U-Net to describe the average DSC of kidney tumors was 85.95% ± 1.46%. 47 The research of Chen et al developed a new cervical cancer segmentation method (called PIC-S-CNN). Chen et al compared this method with 6 different segmentation methods and obtained the best segmentation effect, with an average DSC of 0.84.…”

“…At present, several scholars have applied AI to OAR and CTV for head and neck tumors, lung cancer, breast cancer, prostate cancer, rectal cancer, and cervical cancer. [34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51] Zeineldin et al evaluated the performance of different CNN models in 125 cases of glioma. Compared with manual rendering, the DSC of several CNN models was 81% to 84%.…”