NeuroScan: Medical Image Analysis for Tumor Detection

Abstract—Brain tumors are among the most life-threatening neurological conditions and demand rapid, accurate diagnosis. Traditional diagnosis relies on manual MRI interpretation by radiologists — a process that is slow, inconsistent, and inaccessible in many underserved regions. This paper presents NeuroScan, an automated brain tumor detection system built on convolutional neural network with transfer learning, fine-tuned on 3,000 T1-weighted contrast-enhanced MRI scans. NeuroScan achieves 99.33% test accuracy, 99.34% precision, 99.33% recall, 99.33% F1-score, and 99.86% ROC-AUC on an independent 600-image test set. The preprocessing pipeline includes contour-based intelligent cropping, pixel normalization, resizing to 224×224 pixels, and comprehensive data augmentation (rotation, flipping, zoom, brightness, and shift). The model combines frozen ImageNet-pre-trained ResNet50 layers with custom dense layers, batch normalization, and dropout regularization. Results demonstrate that transfer learning with ResNet50 is an effective and computationally efficient solution for automated radiological brain tumor diagnosis, even when labeled training data is limited. Keywords—Brain Tumor Detection; Deep Learning; ResNet50; Transfer Learning; MRI Image Classification; Convolutional Neural Networks; Medical Image Analysis; Binary Classification; Computer Vision. Introduction Brain tumors represent one of the most serious and complex medical conditions in modern healthcare. They arise from the abnormal, uncontrolled growth of cells within the brain or surrounding tissues, disrupting cognitive function, movement, and sensory perception. Tumors may be benign or malignant; malignant forms such as glioblastoma are often rapidly progressive and carry five-year survival rates as low as 5–15%. According to the American Brain Tumor Association, approximately 700,000 people worldwide receive a primary brain tumor diagnosis every year, with a lifetime risk of roughly 1 in 137 1. Early and accurate detection is therefore directly life-saving. Magnetic Resonance Imaging (MRI) is the gold standard modality for brain tumor evaluation, offering excellent soft-tissue contrast and multi-planar imaging. However, each MRI study consists of 20–30 image slices, and thorough radiologist review typically requires 20–30 minutes per patient. In high-volume clinical settings, this creates serious diagnostic bottlenecks. Inter-observer agreement among radiologists has been reported as low as 85–92% in published studies, meaning diagnostic discrepancy arises in 8–15% of cases 2. In rural and developing regions, severe shortages of specialist neuroradiologists (sometimes 1 per 1–5 million people) cause potentially fatal delays in diagnosis. Deep learning — and specifically Convolutional Neural Networks (CNNs) — has opened transformative new approaches to medical image analysis. CNNs learn discriminative visual features directly from pixel data, without manual feature engineering. Transfer learning, which adapts models pre-trained on large general datasets (such as ImageNet) to specialized domains, makes deep learning practical even when labeled medical training data is scarce 3. This paper presents NeuroScan, a deep learning system for automated binary classification of brain MRI scans as tumor-positive or tumor-negative, achieving 99.33% accuracy through a ResNet50 backbone with transfer learning and comprehensive preprocessing. Related Work A. Traditional Machine Learning Approaches Early Computer-Aided Diagnosis (CAD) systems for brain tumor detection relied on manually engineered features — texture descriptors, shape measurements, intensity statistics, and transform-based features — fed into classifiers such as Support Vector Machines (SVM), k-Nearest Neighbors, and Decision Trees. These systems achieved 85–92% accuracy on research datasets but generalized poorly across different imaging protocols and patient populations, as hand-crafted features rarely captured the full complexity of MRI data 4. B. Deep Learning Approaches The deep learning revolution, catalyzed by AlexNet winning the 2012 ImageNet challenge 5, rapidly transformed medical imaging. CNN architectures such as VGGNet, ResNet, Inception, and DenseNet demonstrated the ability to automatically learn hierarchical feature representations from raw image data, eliminating manual feature engineering entirely. Gupta et al. (2025) applied a modified ResNet50 to 2,577 MRI scans and achieved 97.35% accuracy using ReduceLROnPlateau scheduling 6. Biomedpharma Journal Contributors (2025) conducted a controlled comparison showing ResNet50 (97.91% accuracy, 98.42% AUC) outperforming VGG-19 (95.83%) and custom CNNs (88.6%) on the same dataset, validating residual architectures for this domain 7. Rastogi et al. (2025) proposed a 2D Volumetric CNN achieving 99.3% accuracy on multi-class classification with clinical validation across 500+ patient studies 8. C. Research Gaps Despite impressive accuracy figures, key gaps remain across the published literature: most systems prioritize accuracy over interpretability; few are designed for deployment in resource-limited environments; and many are validated only on small single-institution datasets, raising generalizability concerns. NeuroScan addresses these gaps by combining state-of-the-art accuracy with a lean, efficient architecture suitable for real-world deployment. Dataset and Preprocessing A. Dataset Selection The Figshare Brain Tumor MRI dataset 9 was selected for this study. It contains 3,000 T1-weighted contrast-enhanced MRI images (JPEG format), balanced equally across 1,500 tumor-positive and 1,500 tumor-negative cases, representing 233 unique patients with approximately 13 slices per patient. Images range from 176×224 to 512×512 pixels. All images were annotated by trained radiologists and 95% of the dataset was rated as excellent or good quality. The T1-contrast modality is clinically ideal for tumor detection: gadolinium-based contrast agents accumulate in areas of blood-brain barrier breakdown, making tumors appear as bright enhancing regions — a key visual cue exploited by the model. B. Data Splitting The dataset was divided into three non-overlapping subsets using stratified splitting to preserve the 50-50 class balance in each split: Training set 60% (1,800 images: 900 tumor + 900 normal); Validation set 20% (600 images: 300 + 300); Test set 20% (600 images: 300 + 300). The 60-20-20 split provides sufficient data for transfer learning while ensuring a statistically meaningful evaluation subset. With 600 test samples, the 95% confidence interval for reported accuracy is approximately ±4%. C. Preprocessing Pipeline Raw MRI images undergo three preprocessing steps. 1) Contour-Based Intelligent Cropping: Brain contours are detected using OpenCV findContours. The largest contour's extreme-point bounding box is extracted, removing skull bone, background, and imaging artifacts to focus the model on diagnostically relevant brain parenchyma. 2) Resizing: All cropped images are resized to 224×224 pixels using bilinear interpolation. 3) Normalization: Pixel values are scaled from 0, 255 to 0.0, 1.0 by dividing by 255, consistent with ImageNet pre-training conventions. D. Data Augmentation With only 1,800 training images, data augmentation is essential to prevent overfitting. Keras ImageDataGenerator was configured for on-the-fly training augmentation: rotation ±15°, horizontal and vertical flipping, zoom ±20%, width/height shift ±10%, shearing ±10%, and brightness adjustment ±30%. Augmentation was applied only to the training set, effectively increasing training pattern diversity 5–10×. NeuroScan measured a 5.33 percentage point accuracy improvement from augmentation (~94% baseline vs. 99.33% final), consistent with the 5–8 point gains reported in the medical imaging literature. Proposed Methodology A. Architecture Selection Several architectures were evaluated: VGG-19, Inception-v3, DenseNet, custom CNNs, and ResNet variants. ResNet50 was selected because it achieves 97.91% accuracy (vs. 95.83% for VGG-19) with only 25.6M parameters (vs. 144M for VGG-19 — a 5.6× reduction), inference speed of ~50–70 ms/image, stable training due to residual connections, and it is the most widely cited architecture in medical imaging publications. B. ResNet50 and Residual Learning Traditional deep networks suffer from the vanishing gradient problem: gradients diminish exponentially during backpropagation, making early layers nearly untrainable. He et al. (2016) 3 resolved this with skip connections that allow information to bypass one or more layers: H(x) = F(x) + x, where F(x) is the residual mapping and x is the identity shortcut. Gradients can flow directly to any earlier layer, enabling stable training of 50–152+ layer networks. ResNet50 uses bottleneck blocks: each block applies a 1×1 convolution (dimensionality reduction), a 3×3 convolution (feature extraction), and a 1×1 convolution (dimensionality restoration), dramatically reducing computational cost while preserving representational capacity. C. Transfer Learning Strategy NeuroScan implements a three-stage transfer learning approach. Stage 1 — Load Pre-trained Weights: ResNet50 weights pre-trained on 1.2M ImageNet images (1,000 classes) are loaded, encoding rich representations of edges, textures, shapes, and visual patterns that generalize across domains. Stage 2 — Freeze Base Layers: All 50 ResNet50 convolutional layers are frozen, preserving learned ImageNet features and preventing overfitting on the small MRI dataset. Stage 3 — Custom Classification Head: The head adapts generic ResNet50 features to binary tumor classification: GlobalAveragePooling2D → Dense(512, ReLU) → BatchNorm → Dropout(0.5) → Dense(256, ReLU) → BatchNorm → Dropout(0.3) → Dense(1, Sigmoid). Only the custom head layers are updated