Page 70 - eBook_Proceedings of the International Conference on Digital Manufacturing V2

Proceedings of the International Conference on Digital Manufacturing –
Volume 2

medical contexts, high sensitivity is critical to minimising
missed diagnoses as shown in equation 9.

= (9)
+

3. F1-Score: It is the harmonic mean of precision and recall
(sensitivity). It balances both metrics, offering a single
measure that accounts for the trade-off between the two.
4. Support: It is the number of true ground-truth instances
for each class in the dataset. In other words, it tells you
how many actual samples of that class were present in the
test set
5. Pixel Accuracy per class: It measures the proportion of
correctly classified pixels for a specific class relative to
the total number of ground truth pixels belonging to that
class. It is calculated as shown in equation 10:

= (10)
+

6. Mean Pixel Accuracy: The average of pixel accuracy
values computed independently for each class. It provides
a class-balanced assessment of segmentation performance
by equally weighting all classes, regardless of their
frequency in the dataset. It is calculated as in equation
11:
1
= + ∑ =1 (11)
+

EXPERIMENTS & RESULTS

This section presents a comprehensive evaluation of the proposed
Swin Transformer-based semantic segmentation model through
both quantitative and qualitative analyses. Quantitative results,
summarised in Table 10 and 11, include class-wise recall, F1-
scores, and pixel accuracy metrics, providing measurable insight
into the model’s performance across key cervical cell types. The
confusion matrix (Figure 24), along with bar charts for recall and
pixel accuracy (Figure 25 and 26) further illustrate these patterns.

54