APPLICATION OF MACHINE AND DEEP LEARNING FOR COLORECTAL CANCER EVALUATION IN 3D MRI

Abstract

The aim of this research is twofold. First is related to segmentation of colorectal cancer in 3D MRI, and the second is to characterize the colorectal tumor into two groups; complete responders (CR) and non-responders (NR) to therapy in colorectal cancer. These two studies are conducted in parallel, independently. Study I: Objective: An accurate segmentation of colorectal tumor in 3D magnetic resonance imaging (MRI) volume is an essential requirement in colorectal cancer chemo-radiotherapy. Manual segmentation of colorectal tumor in 3D MRI requires high expertise and subject to laborious work, time consumptions and, inter and intra-observer variability. The primary goal of this research work is to design and develop a straightforward deep learning based algorithm which automatically segments colorectal tumor in 3D T2-weighted (T2w) MRI with reasonable accuracy. Material and Methods: In this study, T2-weighted (T2w) MRI volumes (those were acquired from 43 patients in a sagittal view on a 3.0 Tesla scanner without contrast agent) are used. These patients are diagnosed with a locally advanced colorectal tumor (cT3/T4). In this work, a novel CNN architecture based on a densely connected neural network for volumetric colorectal tumor segmentation is proposed. The proposed CNN architecture contains multi-scale dense inter connectivity between layers of fine and coarse scales, thus by leveraging multi-scale contextual information in the network to get a better flow of information throughout the network. Additionally, the 3D level set algorithm was incorporated as a post-processing task to refine the contours of the network predicted segmentation. Cross-validation was performed in 100 rounds by partitioning the dataset into 30 volumes for training and 13 for testing. Three performance metrics were computed to assess the similarity between predicted segmentation and the true ground truth (i.e., manual segmentation by an expert radiologist/oncologist); including Dice similarity coefficient (DSC), recall rate (RR), and average surface distance (ASD). Results: Above performance metrics were computed in terms of mean and standard deviation (mean ± standard deviation). The DSC, RR, and ASD were (0.84 ± 0.02), (0.85 ± 0.02), and (2.64 ± 2.8) before post-processing; and these performance metrics were (0.86 ± 0.02), (0.87 ± 0.02), and (2.54 ± 2.4) after post-processing, respectively. Conclusion: We compared our proposed method with other existing volumetric medical image segmentation methods (particularly 3D U-net and DenseVoxNet) in our segmentation task. Experimental results reveal that the proposed method has achieved better performance in colorectal tumor segmentation in volumetric MRI than others have. Besides, the proposed method has total parameters approximately 0.7 million due to its simple network architecture, which is much fewer than DenseVoxNet with 1.8 million and 3D U-net with 19.0 million parameters. Study II: Objective: An accurate diagnosis and staging of colorectal cancer at early basis is the supreme interest in the oncology where medical experts have to decide the treatment plan that a patient should go for either therapy or surgical operation. Radiomics is a semiautomatic/automatic quantitative diagnostic technique that decodes the encoded information in large medical imaging datasets, quantitatively. Radiomics measures tumor heterogeneity for diagnosis of several cancers types non-invasively, thus by providing an accurate prognostic or predictive model. Several studies have been carried out to create radiomics based prognostic model for different clinical issues such as patient survival outcome, treatment response, tumor grading, and more where several types of radiomics were used. Therefore, it is difficult to say that what radiomics features are useful in the assessment of colorectal cancer. Hence, the goal of this work is to find which of the radiomics features are the most appropriate in predicting complete tumor response to neoadjuvant therapy, and to assess the possible correlation among these features. Methods: 3D MRI used in this study, was consisted on 43 patients. Consequently, among 43 patients, we have 23 patients observed as complete responders and 20 observed as non-responders. Two different types of radiomics features were extracted from our data; traditional handcrafted radiomics features and deep radiomics features. A total of 109 handcrafted radiomics features were calculated from each MRI volume in this study. Furthermore, 4096 deep radiomics features for each patient, are computed using transfer learning from a pre-trained convolutional neural network (CNN_S). Since high accuracy, efficiency, and reliability are crucial factors in the obtained predictive and prognostic models, which totally depend on the success of radiomics based clinical biomarkers. Therefore, to examine the effectiveness of radiomics based features in achieving an accurate predictive model, it is necessary to validate and compare different machine learning models utilizing all possible radiomics features. For this purpose, in this thesis, the most widely explored supervised machine learning based classifiers were employed. Besides, radiomics have high space dimensionality problem like any high-throughput data-mining field. In this regard, we have assessed the performance of six different feature selection algorithms, which can improve the performance of radiomics based predictive models in different ways. Cross-validation was performed in 100 rounds by partitioning the data as 75% for training and 25% for testing. Results: Using only handcrafted radiomics features, Artificial Neural Network (ANN) classifier and Fisher as feature selection algorithm have achieved the best predictive performance in term of mean area under the ROC curve, AUC, ( i.e., AUCs [mean ± Std]; 0.79 ± 0.016 and 0.8 ± 0.01, respectively). The best prognostic performance using only deep radiomics features was achieved by linear support vector machine (LSVM) classifier and Relief based feature selection algorithm, as 0.8 ± 0.042 and 0.82 ± 0.04, respectively. Whereas, when using a combination of both handcrafted radiomics and deep radiomics features, almost all classifiers in combination with every feature selection algorithm gave better AUC and the best accuracy was given by the LSVM classifier and the Relief based feature selection, as 0.84 ± 0.025 and 0.87 ± 0.013, respectively. Conclusion: we found that the integration of these both handcrafted and deep radiomics features increases the performance of the majority of predictive models. Moreover, the best performance was given by LSVM with all feature selection methods, and Relief based feature selection algorithms gave the best prognostic performance in combination with all classifiers.