Page 63 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Fabrication and Characterization of a Low-Cost Piezoelectric using Rochelle
Salt for Energy Harvesting and Sensor Applications

Additional enhancements like Worst-Case Boosting and
intelligent screening models further improved performance by
minimising false positives (Ming et al., 2022).

Despite their successes, CNNs and hybrid models remain
limited in capturing long-range spatial relationships and
contextual dependencies, especially in large or overlapping
cellular structures. This shortcoming has led to growing interest in
Transformer-based architectures for medical imaging. Initially
developed for natural language processing, transformers utilise
self-attention mechanisms to model global context, making them
especially suitable for high-resolution medical images. Early work
in this direction combined Vision Transformers (ViT) with Class
Activation Mapping (CAM) to identify cancerous regions in
histopathology images without dense pixel-level annotations
(Jahan et al., 2021; Ahmadzadeh Sarhangi, Beigifard, Farmani &
Bolhasani, 2024).

While computationally demanding and less interpretable, this
approach demonstrated the feasibility of weekly supervised
learning. Later, Low-Rank Adaptation (LoRA) was integrated
into ViT frameworks to improve data efficiency and classification
accuracy in cervix-type prediction tasks, outperforming CNN and
ResNet baselines, albeit with concerns regarding model size and
validation transparency (Ghantasala, Hung, Chakrabarti &
Pellakuri, 2024; Hou et al., 2022). Transformers have also proven
valuable in 3D tasks. In (Tanimu, Hamada, Hassan, Kakudi &
Abiodun, 2022), a 3D Transformer encoder and a CNN decoder
outperformed attention-gated U-Nets in radiotherapy dose
prediction. Other innovations like HDC-SAM and CYENET
introduced advanced morphological feature extractors for Pap
smear and colposcopy images, though these systems often relied
on proprietary datasets without external clinical validation (Chitra
& Kumar, 2021).

47