Scaled m-BonacciNet Brings Recurrence-Guided Compression to Medical Image Classification
A new preprint proposes Scaled m-BonacciNet, a lightweight convolutional neural network designed to reduce the computational burden of medical image classification without abandoning accuracy. The work couples a recurrence-inspired mathematical framework with practical CNN engineering, using the dominant m-Bonacci root to rescale channels and depthwise separable convolutions to keep the model compact. In an era where imaging AI is increasingly expected to run outside high-end research servers, the motivation is straightforward: a classifier that performs well but consumes fewer parameters and fewer floating-point operations can be easier to deploy in routine clinical and edge-computing settings.
The paper opens with the construction of the m-Bonacci weighted sequence space, defined for m ≥ 2, 1 ≤ p 0. Within that framework, the authors establish several mathematical properties, including that the space is a Banach space isometrically isomorphic to a related sequence space, with an unconditional Schauder basis. They also describe the space as reflexive and uniformly convex, and they derive a strict inclusion chain, a continuous dual, and an asymptotic truncation estimate. Those details are not decorative. They form the theoretical scaffold for the network design, linking sequence behavior to a channel-scaling strategy intended to preserve representational power while compressing the model.
The engineering idea is compact and easy to state. By scaling the channel structure of a prior Fibonacci-Net architecture with the dominant m-Bonacci root, the model can be tuned across different recurrence orders rather than fixed to a single sequence pattern. Depthwise separable convolutions further reduce cost by splitting spatial filtering from channel mixing, a standard efficiency move that becomes especially valuable in medical imaging, where models often face large inputs and multiple classes. In practice, this kind of architecture seeks to spend parameters only where they matter most, rather than saturating every layer with dense convolutional weight matrices.
The strongest reported performance came from Tribonacci-Net, which achieved 96.09% accuracy and a 0.9581 macro F1-score on a 15-class brain MRI dataset while using 0.85 million parameters and 0.366 GFLOPs. Those numbers matter because macro F1 is more informative than raw accuracy in multi-class settings where class balance may not be uniform and where some diagnostic categories are inherently harder to separate than others. A model that retains a high macro F1-score while remaining small enough to run efficiently has clear appeal for neuroradiology workflows, triage tools, and resource-constrained settings.
A second configuration, Pentanacci-Net, was evaluated on a 44-class dataset and reached 95.87% accuracy. The same model reduced parameters and FLOPs by 47.10% and 44.48%, respectively. Those reductions are clinically relevant because they speak to latency, memory use, and the possibility of broader deployment across hardware that cannot comfortably host heavier vision backbones. For image-heavy specialties, especially where multiple scans may need to be processed rapidly, computational efficiency can influence whether a model stays in the lab or moves closer to the point of care.
Accuracy Gains Came with a Smaller Computational Footprint
A third variant, Hexanacci-Net, was applied to CKD ultrasound and achieved 83.75% accuracy with a 0.7941 macro F1-score. The lower performance relative to the MRI experiments likely reflects the difficulty of the task, the dataset characteristics, or both, but the abstract does not provide the granularity needed to separate those possibilities. What is visible is the broader design pattern: the same recurrence-guided scaling approach was adapted across image domains and class counts, with efficiency preserved as a central constraint rather than an afterthought.
The most important practical point is not the recurrence theory by itself. It is the way that theory is converted into a smaller CNN that still performs competitively on benchmark medical imaging tasks. In a field where model size often balloons as accuracy climbs, an architecture that trims both parameters and FLOPs while remaining within striking distance of stronger baselines can be attractive for real-world integration. Smaller networks are easier to train, easier to store, and easier to replicate on hardware that is common in hospitals and research centers rather than exclusively in large cloud environments.
This preprint remains early-stage research. It has not undergone peer review, and the abstract does not provide detailed information on training protocols, external validation, calibration, subgroup performance, or failure modes. Those omissions matter because imaging classifiers can look strong on benchmark datasets while behaving differently when faced with scanner variation, site-specific artifacts, or cases that fall outside the distribution seen during development. Without that additional testing, the clinical meaning of the reported accuracy and macro F1-scores remains provisional.
The data section indicates that two datasets were used, one publicly available and one available on request, with the brain tumor MRI dataset hosted on Kaggle. The authors also state that the datasets were de-identified and that relevant ethical requirements were followed. Even so, the decisive question for any medical image classifier is not whether it works on a benchmark page, but whether it can tolerate the messiness of real hospital imaging, from protocol differences to noisy inputs and patient-level variation. Until those details are reported in full and the model is validated outside the original experimental setting, the results should be read as a technical advance in efficient image classification rather than a clinical tool ready for deployment.