Brain tumors represent a significant health challenge, with gliomas being the most prevalent primary brain malignancy. Accurate diagnosis, including tumor segmentation, subtype classification, and survival prediction, is crucial for effective treatment planning and prognosis. Traditional machine learning approaches, such as SVMs, KNN, and random forests, often rely on hand-crafted features, limiting their performance. Deep learning offers an advantage by automatically learning optimal features from data. Multimodal MRI (mMRI) provides comprehensive tumor information from different weighted images (T1, T1ce, T2, FLAIR), reflecting phenotypic differences at the cellular level. However, the similar appearance of abnormal tissues in mMRI images makes analysis challenging. This research addresses the need for an integrated framework that simultaneously tackles tumor segmentation, subtype classification, and survival prediction, leveraging the power of deep learning and addressing the challenges posed by mMRI data.

Existing research predominantly focuses on brain tumor segmentation, classification, and survival prediction as independent tasks. While deep learning has shown promise in medical image segmentation and other fields, applying it effectively to brain tumor analysis requires addressing challenges like image quality variation, pre-processing impact, tumor heterogeneity, and data imbalance. Previous studies utilized traditional machine learning methods (SVM, KNN, RF) for brain tumor analysis, but these methods suffered from the limitations of hand-crafted feature extraction. Deep learning methods, such as U-Net and ResNet, have shown superior performance in medical image segmentation, but their application to the integrated problem of segmentation, classification, and survival prediction in brain tumors remains an area of active research and improvement.

The proposed framework integrates three tasks: tumor segmentation, subtype classification, and survival prediction. For segmentation, a novel context-aware deep neural network (CANet) is introduced. CANet incorporates a context encoding module that learns scaling factors to mitigate class imbalance and improve segmentation accuracy. The CANet architecture uses an encoder, context encoding, and decoder structure, similar to U-Net, but enhanced with the context encoding module to capture global context. The loss function combines Dice loss and semantic loss (*L* = *L_dice* + *L_se*). The segmented tumor regions (enhancing tumor (ET), tumor core (TC), whole tumor (WT)) are then used as input for the 3D CNN-based tumor subtype classification. Finally, a hybrid approach for survival prediction uses the CANet's encoding module for feature extraction, incorporates patient age, employs LASSO feature selection, and utilizes linear regression to predict survival days. The datasets used are the BraTS 2019 and CPM-RadPath 2019 challenges’ datasets. Data augmentation (rotation, scaling, cropping) was used to mitigate data imbalance and limited training data size. Evaluation metrics included Dice similarity coefficient (DSC), Hausdorff distance (HD95), classification accuracy, and mean squared error (MSE).

The CANet achieved superior performance in brain tumor segmentation compared to ResNet, UNet, and UNet-VAE on the BraTS 2019 validation dataset. The average DSC values on the testing dataset were 0.821 for ET, 0.895 for WT, and 0.835 for TC. The HD95 values were 3.319 mm for ET, 4.897 mm for WT, and 6.712 mm for TC. In the CPM-RadPath 2019 challenge, the proposed tumor classification method achieved second place in the testing phase with a DSC of 0.639. The survival prediction model achieved a testing accuracy of 0.484 and an MSE of 334,492 on the BraTS 2019 dataset. The study also compared the proposed survival prediction method with a traditional machine learning approach, showing that the proposed method significantly outperformed the conventional method in terms of accuracy and MSE. Analysis of age and gender impact on survival indicated some trends, but statistical analysis (ANOVA) did not show significant effects, likely due to a limited sample size.

The integrated framework presented in this study successfully addresses the challenges of brain tumor analysis by jointly performing segmentation, classification, and survival prediction. The proposed CANet demonstrates significant improvements in segmentation accuracy compared to existing architectures, highlighting the effectiveness of the context encoding module in mitigating class imbalance and capturing global contextual information. The strong performance in the CPM-RadPath 2019 challenge underscores the robustness and generalizability of the proposed classification method. Although the survival prediction accuracy is moderate, the integration with the segmentation and classification tasks offers a unique perspective on patient prognosis. The limitations of the study, including data imbalance and potential biases in the datasets, need consideration when interpreting the results. The integration of various data sources, like BraTS 2019, BraTS 2020, and TCIA, aimed to improve robustness and generalizability.

This research presents a novel integrated framework for brain tumor analysis using context-aware deep learning. The proposed CANet achieved state-of-the-art performance in tumor segmentation and subtype classification, while the hybrid survival prediction method yielded promising results. Future work could focus on incorporating additional data modalities (e.g., pathology images, molecular genetic data) to enhance classification accuracy and survival prediction, aligning with the updated WHO classification criteria. Addressing the data imbalance issue through advanced sampling techniques is also important.

The study acknowledges several limitations: data imbalance in both segmentation (edema samples more prevalent) and survival prediction (uneven distribution across time categories) could affect model performance. The sample size, particularly for the survival prediction task, may not be large enough for conclusive statistical analysis on the impact of gender and age. The reliance on publicly available datasets may also introduce biases that limit the generalizability of findings to diverse patient populations.