Evaluation of Lightweight CNN Architectures for Adrenal Cortical Carcinoma Classification from MRI: A Comparative Study of Pre-trained and Custom Models
Sachchit Balamurugan, Avdhoot Datar
Texas Online Preparatory School, Lewisville, USA
Publication date: November 20, 2025
Texas Online Preparatory School, Lewisville, USA
Publication date: November 20, 2025
DOI: http://doi.org/10.34614/JIYRC2025II23
ABSTRACT
Accurate classification of adrenal cortical carcinoma (ACC) using magnetic resonance imaging (MRI) remains a key challenge in medical image analysis. In recent years, convolutional neural networks (CNNs) have been developed to analyze medical images. The study addresses two primary research questions: (a) Do pre-trained CNNs outperform custom-trained models in ACC classification from MRI? and (b) Does a lower parameter count correspond to improved predictive efficiency and generalization? Four pre-trained 6 CNN architectures (VGG16, ResNet50, MobileNetV2, ConvNeXtTiny) and two custom models optimized using Optuna hyperparameter tuning were evaluated on a dataset comprising MRI scans of ACC and normal adrenal tissue. The dataset was divided into training, validation, and test subsets. Classification performance was assessed using accuracy, precision, recall, specificity, confusion matrices, and receiver operating characteristic (ROC) analysis. MobileNetV2 demonstrated the strongest overall performance, achieving the highest test recall (0.95), specificity (0.99), and AUC (0.96), despite having only 3.57 million parameters. Both MobileNetV2 and ResNet50 achieved the highest test accuracy (92.2%). While custom models underperformed overall, the smallest custom model (Optuna Model 2, 1.16M parameters) outperformed the larger custom model in test accuracy (85.5% vs. 80.3%) and total misclassifications. ROC analysis further confirmed the effectiveness of smaller, well-optimized models (AUC = 0.93 for Optuna Model 2). The results confirm that while pre-trained models offer high classification performance, compact CNNs with optimized architectures can deliver comparably strong results with drastically fewer parameters. These findings emphasize the potential of lightweight deep learning models for robust and resource-efficient medical image classification.
Accurate classification of adrenal cortical carcinoma (ACC) using magnetic resonance imaging (MRI) remains a key challenge in medical image analysis. In recent years, convolutional neural networks (CNNs) have been developed to analyze medical images. The study addresses two primary research questions: (a) Do pre-trained CNNs outperform custom-trained models in ACC classification from MRI? and (b) Does a lower parameter count correspond to improved predictive efficiency and generalization? Four pre-trained 6 CNN architectures (VGG16, ResNet50, MobileNetV2, ConvNeXtTiny) and two custom models optimized using Optuna hyperparameter tuning were evaluated on a dataset comprising MRI scans of ACC and normal adrenal tissue. The dataset was divided into training, validation, and test subsets. Classification performance was assessed using accuracy, precision, recall, specificity, confusion matrices, and receiver operating characteristic (ROC) analysis. MobileNetV2 demonstrated the strongest overall performance, achieving the highest test recall (0.95), specificity (0.99), and AUC (0.96), despite having only 3.57 million parameters. Both MobileNetV2 and ResNet50 achieved the highest test accuracy (92.2%). While custom models underperformed overall, the smallest custom model (Optuna Model 2, 1.16M parameters) outperformed the larger custom model in test accuracy (85.5% vs. 80.3%) and total misclassifications. ROC analysis further confirmed the effectiveness of smaller, well-optimized models (AUC = 0.93 for Optuna Model 2). The results confirm that while pre-trained models offer high classification performance, compact CNNs with optimized architectures can deliver comparably strong results with drastically fewer parameters. These findings emphasize the potential of lightweight deep learning models for robust and resource-efficient medical image classification.
