Sandi Baressi Šegota1*, Ivan Lorencin1, Domagoj Frank2 and Nikola Anđelić3
1Faculty of Informatics Pula – Juraj Dobrila University of Pula
2Department of Computer Science and Informatics, University North
3Faculty of Engineering – University of Rijeka
sandi.baressi.segota [at] unipu.hr
Abstract
Accurate classification of medical imagery presented a critical challenge in modern diagnostic workflows, particularly when operating under constrained computational resources. This study addressed the need for high-performance, lightweight diagnostic tools by developing and evaluating an ensemble of small-scale Convolutional Neural Networks (CNNs). The primary objective aimed to demonstrate that a confidence-based combination of computationally undemanding architectures achieved robust classification accuracy comparable to larger, resource-intensive models.
Researchers utilized the OrganAMNIST dataset from the widely standardized MedMNIST collection, a comprehensive repository of abdominal CT images classified into eleven distinct organ categories. The methodology involved designing, training, and validating five unique, lightweight CNN architectures. These models included a basic three-layer SimpleCNN, a reduced VGG8, a compact ResNet10, a custom InceptionMini, and a TinyCNN. Investigators trained each base model independently from scratch for ten epochs using the Adam optimizer and cross-entropy loss. Following individual model optimization, the study implemented a confidence-based ensemble technique. This approach aggregated the predictive probabilities from all five architectures, and it weighted the ultimate classification decision based on the varying confidence levels each model exhibited for specific classes.
Our empirical evaluation revealed significant individual performance among the baseline models. During the validation phase, individual architectures consistently achieved high predictive accuracy, with initial basic models like SimpleCNN and VGG8 surpassing 96% and 97% accuracy, respectively. Subsequently, the confidence-based ensemble method effectively synthesized these individual strengths. By dynamically leveraging the highest confidence predictions across the varied feature-extraction paradigms of the five subset models, the ensemble framework reduced generalization errors and improved overall predictive stability across all eleven target classes.
In conclusion, this research validated the efficacy of employing a confidence-weighted ensemble of minimized CNN architectures for complex medical image classification tasks. The implemented methodology successfully mitigated the typical trade-off between computational efficiency and diagnostic precision. The findings demonstrated that intelligently combined lightweight neural networks offered a highly scalable, accurate, and resource-efficient solution for automated abdominal CT scan analysis, which facilitated broader deployment of artificial intelligence in diverse clinical environments.
Keywords: CNN, Ensemble, Performance Testing
Acknowledgement: This work was (partially) supported by the EC Digital Europe Programme EDIH Adria 2.0 (101256325); SPIN projects IP.1.1.03.0120, IP.1.1.03.0158 and IP.1.1.03.0039; and NextGenerationEU University grants: uniri-iz-25-6, uniri-iz-25-220, IIP_010144, IIP_010136, and UNIN-TEH-25-1-8.