Abstract
<jats:p>Cervical cancer screening remains highly unequal across regions, with low and middle income countries facing limited access to expert cytology and laboratory infrastructure. This work frames low-cost cervical cytology screening as an intelligent healthcare environment in which image acquisition, quality enhancement, embedded AI inference, and clinician-oriented decision support are combined at the point of care. We present the AI core of such an environment, designed for integration into a low-cost digital cytology acquisition station operating under heterogeneous imaging conditions and limited connectivity. First, we model realistic quality variability by generating datasets with multiple degradation profiles and apply a restoration stage based on Real-ESRGAN to enhance low-quality inputs. Second, we train a compact convolutional neural network whose hyperparameters are selected via Bayesian optimization, aiming at an accuracy-efficiency trade-off suitable for embedded edge inference. Finally, we apply TinyML-oriented compression, combining pruning and INT8 quantization with export to Tensor-Flow Lite. On an internal held-out test split, the baseline model achieved 0.888 accuracy, while the mixed-quality enhanced configuration reached 0.922, indicating improved robustness under more realistic acquisition conditions. Model compression reduced the footprint from 21.18 MB to 2.74 MB, supporting feasibility for deployment on constrained hardware. Overall, the results suggest that quality-aware training, image enhancement, and TinyML compression can enable privacy-preserving, offline, and scalable intelligent point-of-care environments for cervical cytology triage in resource-constrained clinical settings.</jats:p>