Page 76 - AIH-2-3

Artificial Intelligence in Health CNN model for leukemia diagnosis

Table 2. Conceptual representation of models’ accuracy rates change over epochs
Epochs Model Accuracy Precision Recall F1‑Score Method for handling Key contributions
(%) (%) (%) (%) imbalanced data
10 Customized DL Model 85 82 83 82.5 Tversky loss function High accuracy, robust to
imbalanced datasets
10 ResNet101 Ensemble 75 72 73 72.5 No explicit method Performs well but not
optimized for imbalances
10 ALLNET Model 88 85 86 85.5 Not specified High accuracy, good at
distinguishing subtypes
10 SVM + CNN Hybrid 70 68 70 69 No explicit method Basic hybrid model, limited
by data imbalance
10 Traditional ML 60 58 60 59 No explicit method Struggles with imbalanced
datasets
20 Customized DL Model 90 88 89 88.5 Tversky loss function Improved handling of subtle
variations in cells
20 ResNet101 Ensemble 80 78 79 78.5 No explicit method Improved but imbalances are
not addressed
20 ALLNET Model 90 88 89 88.5 Not specified Strong performance, close to
customized DL model
20 SVM + CNN Hybrid 75 72 73 72.5 No explicit method Some improvement but
limited by data imbalance
20 Traditional ML 65 63 65 64 No explicit method Slight improvement but still
lacking robustness
30 Customized DL Model 93 91 92 91.5 Tversky loss function Outstanding performance,
very effective in handling
imbalances
30 ResNet101 Ensemble 82 80 81 80.5 No explicit method Stable but still weaker at
handling imbalances
30 ALLNET Model 92 89 90 89.5 Not specified Continues to perform well
across metrics
30 SVM + CNN Hybrid 78 75 76 75.5 No explicit method Better results but still limited
by hybrid model
30 Traditional ML 68 66 68 67 No explicit method Performance improvement
but still behind DL models
40 Customized DL Model 96 94 95 94.5 Tversky loss function Peak performance in
accuracy and handling
imbalances
40 ResNet101 Ensemble 82 80 81 80.5 No explicit method Plateauing performance, not
addressing imbalances well
40 ALLNET Model 94 92 93 92.5 Not specified Continues to perform
exceptionally well
40 SVM + CNN Hybrid 80 77 78 77.5 No explicit method Gradual improvements but
limited by imbalance
40 Traditional ML 70 68 70 69 No explicit method Steady performance, still
underperforms DL models
50 Customized DL Model 99 98 98.5 98 Tversky loss function Nearly perfect accuracy and
robustness to data imbalance
50 ResNet101 Ensemble 82 80 81 80.5 No explicit method Performance stabilizes, still
not handling imbalances well
50 ALLNET Model 95.54 94 94.5 94.25 Not specified Reaches top-tier performance
but behind customized DL
model
(Cont’d...)

Volume 2 Issue 3 (2025) 70 doi: 10.36922/aih.4710