Page 62 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Proceedings of the International Conference on Digital Manufacturing –
Volume 2

for segmentation, and SqueezeNet for feature extraction,
achieving a 97.96% classification accuracy.

Beyond individual networks, hybrid and ensemble-based
frameworks began to dominate the field. Works such as (Zak et
al., 2022) and (Alyafeai & Ghouti, 2020) introduced bagging
classifiers and multi-scale pyramid features using CNNs like
DarkNet19 and DarkNet53, followed by Neighborhood
Component Analysis (Huang, Yang, Li, He & Liang, 2021) and
SVM classification. Other studies leveraged fuzzy learning
ensembles (AlMohimeed, Saleh, Mostafa, Saad & Talaat, 2023)
and (Fang, Lei, Liao & Wu, 2022) and combined CNN outputs
from pre-trained models like ResNet-50, ResNet-101, and
GoogleNet to improve diagnostic robustness via voting strategies
(Wang et al., 2024). In parallel, attention was given to analysing
whole-slide images (WSIs), moving beyond cropped cell images.
Studies (Ding et al., 2021) and (Kruczkowski et. al., 2022)
proposed modular architectures consisting of segmentation
blocks, compact CNN variants (e.g., Compact-VGG), and
recurrent models to extract lesion-level probabilities from WSIs.
Mask R-CNN was also used for precise instance-level
segmentation, followed by secondary classification using VGG-
style networks (Lilhore et al., 2022). Meanwhile, integrating
mobile devices and Internet-of-Things (IoT) systems, as
demonstrated by the μSmartScope-enabled diagnostic system,
showcased the potential of lightweight deep learning models in
low-resource environments (Tan et al., 2021). Advanced object
detection networks like Faster R-CNN, YOLOv3, and Tiny
YOLO were benchmarked, with YOLOv3 (608 input size)
achieving the highest mean Average Precision (mAP = 0.574)
(Ontor et al., 2022). These detection-centric approaches marked a
shift toward real-time and scalable diagnostic solutions. Graph
Convolutional Networks (GCNs) have recently been applied to
cytology tasks. Studies (Shandilya, Anand, Chauhan, Pokhariya
& Gupta, 2024) and (Hong, Xiong, Yang & Mo, 2024) used GCNs
to map structural relationships between cells, enabling the
classification of diverse cell types such as parabasal, dyskeratotic,
and koilocytotic.

46