An Enhanced Brain Tumor Classification using Enhanced Squeeze and Excitation Network with Long Short-Term Memory

Context & Rationale

Brain tumour diagnosis presents one of the most formidable challenges in clinical oncology, with five-year survival rates substantially lower than those observed in many other malignancies. Accurate and timely classification of brain tumours from magnetic resonance imaging (MRI) is critical for treatment planning, surgical intervention, and prognosis assessment; yet conventional diagnostic pathways ??? including biopsy ??? are typically conducted only after definitive surgical procedures, introducing significant delay and risk. The integration of machine learning into radiological workflows therefore represents a clinically consequential objective, with Convolutional Neural Networks (CNNs) having emerged as the principal paradigm for automated image classification tasks of this nature.

Scope of the Study

This study presents the design and implementation of a novel hybrid deep learning architecture termed SE-LSTM, which integrates an Enhanced Squeeze-and-Excitation (SE) Network with a Long Short-Term Memory (LSTM) recurrent neural network. The Squeeze-and-Excitation component operates through a channel-wise recalibration mechanism ??? employing global average pooling for feature aggregation (squeeze) and a self-gating function for adaptive feature reweighting (excitation) ??? thereby enabling the network to selectively emphasise informative feature channels. The LSTM component extends this capability to the temporal and sequential modelling of high-dimensional feature representations, addressing optimisation challenges that arise in large-scale brain tumour datasets.

Architecture & Experimental Design

The proposed pipeline encompasses a structured sequence of processing stages: image resizing and normalisation, Wiener filter-based denoising, batch normalisation, SE network feature extraction, LSTM-based sequential learning, and final softmax classification. The model was evaluated on a brain tumour MRI dataset in MATLAB simulation, with classification performance assessed through sensitivity, specificity, and accuracy metrics derived from a binary confusion matrix framework. Comparative benchmarking against established architectures ??? including Genetic Algorithm-enhanced CNN (GA-CNN) and a standard deep learning baseline ??? was undertaken to contextualise the contribution of the proposed method.

Key Findings

The SE-LSTM architecture demonstrated a statistically meaningful improvement in classification accuracy relative to both comparator methods across two independent study conditions. The magnitude of this improvement, and the specific accuracy values recorded for each method and study configuration, are reported in full within the publication and provide a basis for evaluating the practical utility of the proposed architecture in clinical diagnostic support applications.

Keywords

Brain tumour classification  ??  MRI  ??  Convolutional neural network  ??  Squeeze-and-Excitation network  ??  Long Short-Term Memory  ??  Deep learning  ??  Image segmentation  ??  SE-LSTM  ??  GA-CNN  ??  Medical image analysis