Robust Brain Tumor Diagnosis in Clinical MRI Scenarios: A Non-Negative Matrix Tri-Factorization Approach for Missing Modality Imputation

Brain tumors are among the most common and aggressive diseases, often resulting in significantly reduced life expectancy at advanced stages. Consequently, effective treatment planning becomes crucial for improving patient quality of life. However, in real-world clinical settings, MRI data often suffer from missing modalities i.e., certain MRI sequences (such as T1, T2, or FLAIR) are unavailable due to patient movement, time constraints, or equipment failure. This partial data availability leads to reduced performance of deep learning-based diagnostic systems, which typically rely on input. To address this gap, we propose a novel framework based on Non-negative Matrix Tri-Factorization (NMTF) that reconstructs missing portions of MRI data and enables reliable classification even when input is incomplete. Unlike previous studies that either ignore missing data or assume fixed missing patterns, our model adaptively factorizes available information into three low-rank non-negative matrices, enabling the recovery of missing features. To evaluate the effectiveness of proposed framework, six deep neural network classifiers are initially trained and tested on complete MRI scans (both 2D and 3D). Subsequently, varying amounts of data (from 10% to 50%) from sagittal slice are deliberately removed, and the classifiers are re-applied, resulting in a noticeable drop in performance due to the missing information. To address this, the proposed NMTF method is used to reconstruct the missing portions of the MRI data. The classifiers are then re-evaluated on the NMTF-recovered scans. This approach is validated across two distinct datasets. Results indicate that the average classification accuracy improves by approximately 10% ± 2 when comparing performance on incomplete MRI scans versus those reconstructed via NMTF.