An end-to-end agentic framework compares CNNs and vision transformers across three medical imaging tasks and adds LLM-generated structured reports

The framework performs an end-to-end comparative evaluation of ResNet50, EfficientNet-B0, DenseNet121, DeiT-Small, and Swin-Tiny across three heterogeneous medical imaging tasks: chest X-ray pneumonia classification, brain MRI tumor detection, and dermoscopic skin cancer classification. Compared with common practice that focuses on a single task or a single architecture, this work places multiple CNNs and vision transformers within one evaluation pipeline for cross-dataset comparison. Evidence comes from the authors' description of experiments across three tasks and five architectures; the loaded text is incomplete and does not report dataset sizes, split procedures, or numerical metrics.

The framework integrates transfer learning, class-imbalance handling, model calibration, bootstrap confidence intervals, robustness evaluation, failure-case analysis, Grad-CAM explainability, ensemble learning, and extensive performance metrics into a single pipeline. Compared with reporting predictive accuracy alone, this work treats confidence estimation, perturbation robustness, and interpretability as integral parts of the system. Evidence comes from the authors' enumeration of framework components; the loaded text does not provide the specific experimental settings or result values for each component.

The framework adds an LLM-driven reporting component that generates structured model-comparison summaries, explainability interpretations, and decision-support reports, and is constrained to research-oriented assistance rather than autonomous clinical diagnosis. Compared with using an LLM directly for diagnostic output, this work positions the LLM at the level of report generation and decision support while explicitly stating its non-autonomous-diagnosis boundary. Evidence comes from the authors' description of the LLM layer's purpose and constraint; the loaded text does not give the model used, prompt design, or report-quality evaluation details.

Experimental results indicate that CNN architectures remain highly effective for structured radiographic images, while vision transformers provide strong performance and robustness for the more contextually complex brain MRI and dermoscopic imaging tasks. This conclusion links architecture choice to how structured and contextually complex the imaging task is, rather than declaring one architecture universally superior. Evidence comes from the authors' summary statement of experimental results; the loaded text does not provide specific metrics, confidence intervals, or robustness perturbation values.

The framework targets research and comparative evaluation for three medical image classification tasks: chest X-ray pneumonia classification, brain MRI tumor detection, and dermoscopic skin cancer classification. It is suited to readers who want to examine predictive performance, calibration, robustness, explainability, and report generation within one pipeline. The LLM layer is constrained to research-oriented assistance for generating model-comparison summaries, explainability interpretations, and decision-support reports rather than autonomous clinical diagnosis, so its intended setting is research evaluation and decision support, not replacement of clinical judgment.

The loaded text is incomplete and lacks dataset sizes, split procedures, specific performance metrics, calibration and robustness values, and evaluation details for LLM report quality, so the robustness of the conclusions cannot be judged. Readers may still watch: whether the CNN versus vision transformer difference is stable across tasks; how much class-imbalance handling and ensemble learning affect results; which perturbations the robustness evaluation uses; and how LLM-generated reports connect to and are checked against clinical decision support.