Improved content-based brain tumor retrieval for magnetic resonance images using weight initialization framework with densely connected deep neural network

Content-based medical image retrieval (CBMIR) systems can assist and help doctors and radiologists in reliable diagnosis by retrieving relevant medical cases from past cases that share similarities with the current case. Therefore, there is a need for improved retrieval accuracy in CBMIR systems. Retrieving similar Brain Tumor magnetic resonance imaging (MRI) slices from the same class as the query is inherently challenging. For different types of tumors, there is no class-specific structure, size or shape. Further, MRIs being present in multiple views increase discrepancy in interview retrieval. This leads to high inter-class similarity but at the same time high intra-class variations. Skewed data quantities for Brain Tumor types like Meningioma further add to another challenge. It is necessary to formulate rich and generic MRI representations to address these issues. Toward the same, it is essential to model spatial contexts on a multi-scale across the tumor-affected regions and enhances the feature representational learning by extracting generic features. Hence, we propose a Weight Initialization Framework with Densely Connected Networks to improve generalization for Brain Tumor MRI retrieval. The proposed framework uplifts DenseNet-based models for feature extraction as they incorporate feature reuse and feature learning in a multi-scale manner. Further, a weight Initialization Framework (WIF) is used for improvising the representational learning. Specifically, WIF initializes weights of the DenseNet model by transfer learning-based adaptation, which then involves freezing the initial few layers. The freezing step ensures rich low-level features, even for long-tailed classes like Meningioma, while the remaining trainable layers are fine-tuned to incorporate domain-inherent feature learning. The proposed approach outperforms state-of-the-art by a margin of 1.70% and 1.69% on standard mAP and p@10, respectively. Concretely, when DenseNet and WIF are jointly employed, a stark increment in performance for the Meningioma class is observed, suggesting the generalizability of the framework.