Depeursinge, A. | Search bioRxiv

bioRxiv2024

Exploiting XAI maps to improve MS lesion segmentation and detection in MRI

To date, several methods have been developed to explain deep learning algorithms for classification tasks. Recently, an adaptation of two of such methods has been proposed to generate instance-level explainable maps in a semantic segmentation scenario, such as multiple sclerosis (MS) lesion segmentation. In the mentioned work, a 3D U-Net was trained and tested for MS lesion segmentation, yielding an F1 score of 0.7006, and a positive predictive value (PPV) of 0.6265. The distribution of values in explainable maps exposed some differences between maps of true and false positive (TP/FP) examples. Inspired by those results, we explore in this paper the use of characteristics of lesion-specific saliency maps to refine segmentation and detection scores. We generate around 21000 maps from as many TP/FP lesions in a batch of 72 patients (training set) and 4868 from the 37 patients in the test set. 93 radiomic features extracted from the first set of maps were used to train a logistic regression model and classify TP versus FP. On the test set, F1 score and PPV were improved by a large margin when compared to the initial model, reaching 0.7450 and 0.7817, with 95% confidence intervals of [0.7358, 0.7547] and [0.7679, 0.7962], respectively. These results suggest that saliency maps can be used to refine prediction scores, boosting a models performances.

bioengineering↗

bioRxiv2018

Locoregional Radiogenomic Models Capture Gene Expression Heterogeneity in Glioblastoma

Radiogenomics mapping noninvasively determines important relationships between the molecular genotype and imaging phenotype of various tumors, allowing advances in both clinical care and cancer research. While early work has shown its technical feasibility, here we extend radiogenomic mapping to a locoregional level that can account for the molecular heterogeneity of tumors. To achieve this, our data processing pipeline relies on three main steps: 1) the use of multi-omics data fusion to generate a set of 100 interpretable gene modules, 2) the use of patch-based image analysis (specifically of contrast-enhanced T1-weighted weighted MR images) combined with Generalized Linear Models (GLM) to establish potential links between module expressions and local MR signal, and 3) the use of expression heatmaps based on GLMs decision values to explore visualization of tumor molecular heterogeneity. The performance of the proposed approach was evaluated using a leave-one-patient-out crossvalidation method as well as a separate validation data set. The top performing models were based on a small set of 20 features and yielded Area Under the receiver operating characteristic Curve (AUC) above 0.65 on the validation cohort for eight modules. Next, we demonstrate the clinical and biological interpretation of four modules using molecular expression heatmaps superimposed on clinical radiographic images, showing the potential for assessing tumor molecular heterogeneity and the utility of this method for precision treatment in clinical decision making and imaging surveillance.

bioinformatics↗