Introduction
Related Studies
Materials and Methods
Dataset Preparation
ROI Extraction
Data Augmentation
Dataset Bifurcation
DDSM dataset | Mini-MIAS | ||
---|---|---|---|
Total cases | 4480 | 322 | |
Total ROIs | 4880 | 322 | |
Augmented sample | 4880 × 10 = 48800 ROIs | 322 × 20 = 6440 ROIs | |
Dataset bifurcation | Balanced bifurcation | Training set: 24400 | Training set: 3220 |
Testing set: 24400 | Testing set: 3220 | ||
Unbalanced bifurcation | Training set: 34160 | Training set: 4508 | |
Testing set: 14640 | Testing set: 1932 |
Proposed Work
Deep Learning Model
Convolutional Layer
Activation Layer
Pooling Layer
AlexNet
ResNet
Stochastic Gradient Descent
Performance Evolutionary Parameters
Kappa Coefficient
Experiments and Results
Experimental Setup
Experiments
MIAS ROIs | DDSM ROIs | |
---|---|---|
Balanced Bifurcation | Training set: 3220 | Training set: 24400 |
Testing set: 3220 | Testing set: 24400 | |
Unbalanced Bifurcation | Training set: 4508 | Training set: 34160 |
Testing set: 1932 | Testing set: 14640 |
Dataset | Dataset bifurcation | Experiment no | Description |
---|---|---|---|
MIAS | Balanced bifurcation | Experiment 1 | Dense tissue pattern characterization using AlexNet and ResNet-18 model |
Unbalanced bifurcation | Experiment 2 | Dense tissue pattern characterization using AlexNet and ResNet-18 model | |
DDSM | Balanced bifurcation | Experiment 3 | Dense tissue pattern characterization using AlexNet and ResNet-18 model |
Unbalanced bifurcation | Experiment 4 | Dense tissue pattern characterization using AlexNet and ResNet-18 model | |
DDSM + MIAS | Balanced bifurcation | Experiment 5 | Dense tissue pattern characterization using AlexNet and ResNet-18 model |
Unbalanced bifurcation | Experiment 6 | Dense tissue pattern characterization using AlexNet and ResNet-18 model |
Experiment 1
DL model | Activation function | Confusion matrix | Accuracy (%) | Kappa coefficient (\({\boldsymbol{\tau}}\)) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 897 | 163 | 88.9 | 0.750 |
Dense | 194 | 1966 | ||||
Sigmoid | Fatty | 897 | 163 | 86.3 | 0.698 | |
Dense | 278 | 1882 | ||||
Tanh | Fatty | 895 | 165 | 86.1 | 0.694 | |
Dense | 282 | 1878 | ||||
Leaky ReLU | Fatty | 873 | 187 | 85.0 | 0.669 | |
Dense | 296 | 1864 | ||||
ResNet-18 | ReLU | Fatty | 908 | 152 | 89.8 | 0.770 |
Dense | 176 | 1984 | ||||
Sigmoid | Fatty | 864 | 196 | 87.2 | 0.711 | |
Dense | 216 | 1944 | ||||
Tanh | Fatty | 870 | 190 | 85.0 | 0.668 | |
Dense | 293 | 1867 | ||||
Leaky ReLU | Fatty | 873 | 187 | 85.2 | 0.673 | |
Dense | 289 | 1871 |
Experiment 2
DL model | Activation function | Confusion matrix | Accuracy (%) | \({\boldsymbol{\tau}}\) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 569 | 67 | 89.3 | 0.766 |
Dense | 138 | 1158 | ||||
Sigmoid | Fatty | 581 | 55 | 89.1 | 0.763 | |
Dense | 155 | 1141 | ||||
Tanh | Fatty | 548 | 88 | 86.1 | 0.698 | |
Dense | 179 | 1117 | ||||
Leaky ReLU | Fatty | 541 | 91 | 88.6 | 0.749 | |
Dense | 125 | 1171 | ||||
ResNet-18 | ReLU | Fatty | 581 | 55 | 91.3 | 0.807 |
Dense | 113 | 1183 | ||||
Sigmoid | Fatty | 570 | 66 | 89.6 | 0.770 | |
Dense | 135 | 1158 | ||||
Tanh | Fatty | 568 | 68 | 87.8 | 0.734 | |
Dense | 168 | 1128 | ||||
Leaky ReLU | Fatty | 558 | 78 | 86.5 | 0.706 | |
Dense | 183 | 1113 |
Experiment 3
DL model | Activation function | Confusion matrix | Accuracy (%) | \({\boldsymbol{\tau}}\) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 10849 | 1451 | 88.2 | 0.764 |
Dense | 1428 | 10672 | ||||
Sigmoid | Fatty | 11123 | 1177 | 86.3 | 0.725 | |
Dense | 2166 | 9934 | ||||
Tanh | Fatty | 10418 | 1882 | 85.6 | 0.712 | |
Dense | 1631 | 10469 | ||||
Leaky ReLU | Fatty | 10568 | 1732 | 86.0 | 0.720 | |
Dense | 1684 | 10416 | ||||
ResNet-18 | ReLU | Fatty | 11112 | 1188 | 90.2 | 0.803 |
Dense | 1206 | 10894 | ||||
Sigmoid | Fatty | 10910 | 1390 | 88.7 | 0.773 | |
Dense | 1368 | 10732 | ||||
Tanh | Fatty | 10676 | 1624 | 86.8 | 0.736 | |
Dense | 1597 | 10503 | ||||
Leaky ReLU | Fatty | 11218 | 1082 | 89.7 | 0.794 | |
Dense | 1420 | 10680 |
Experiment 4
DL model | Activation function | Confusion matrix | Accuracy (%) | \({\boldsymbol{\tau}}\) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 6389 | 991 | 87.8 | 0.756 |
Dense | 794 | 6466 | ||||
Sigmoid | Fatty | 6332 | 1048 | 85.1 | 0.702 | |
Dense | 1133 | 6127 | ||||
Tanh | Fatty | 6103 | 1277 | 84.6 | 0.692 | |
Dense | 977 | 6283 | ||||
Leaky ReLU | Fatty | 6456 | 924 | 87.1 | 0.742 | |
Dense | 964 | 6296 | ||||
ResNet-18 | ReLU | Fatty | 6812 | 568 | 92.3 | 0.846 |
Dense | 559 | 6701 | ||||
Sigmoid | Fatty | 6718 | 662 | 89.0 | 0.780 | |
Dense | 946 | 6314 | ||||
Tanh | Fatty | 6287 | 1093 | 87.3 | 0.746 | |
Dense | 763 | 6497 | ||||
Leaky ReLU | Fatty | 6676 | 704 | 90.3 | 0.807 | |
Dense | 702 | 6558 |
Experiment 5
DL model | Activation function | Confusion matrix | Accuracy (%) | \({\boldsymbol{\tau}}\) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 11583 | 1777 | 86.7 | 0.733 |
Dense | 1897 | 12363 | ||||
Sigmoid | Fatty | 11336 | 2024 | 85.6 | 0.730 | |
Dense | 1654 | 12306 | ||||
Tanh | Fatty | 11429 | 1931 | 84.8 | 0.695 | |
Dense | 2268 | 11992 | ||||
Leaky ReLU | Fatty | 11649 | 1711 | 87.2 | 0.743 | |
Dense | 1825 | 12435 | ||||
ResNet-18 | ReLU | Fatty | 12097 | 1263 | 88.3 | 0.766 |
Dense | 1969 | 12291 | ||||
Sigmoid | Fatty | 11730 | 1630 | 87.8 | 0.755 | |
Dense | 1740 | 12520 | ||||
Tanh | Fatty | 11416 | 1944 | 86.2 | 0.723 | |
Dense | 1868 | 12392 | ||||
Leaky ReLU | Fatty | 11667 | 1693 | 88.1 | 0.761 | |
Dense | 1594 | 12666 |
Experiment 6
DL model | Activation function | Confusion matrix | Accuracy (%) | \({\boldsymbol{\tau}}\) | ||
---|---|---|---|---|---|---|
Fatty | Dense | |||||
AlexNet | ReLU | Fatty | 7378 | 638 | 90.8 | 0.815 |
Dense | 888 | 7668 | ||||
Sigmoid | Fatty | 7014 | 1002 | 87.5 | 0.749 | |
Dense | 1070 | 7486 | ||||
Tanh | Fatty | 6965 | 1051 | 86.9 | 0.736 | |
Dense | 1130 | 7426 | ||||
Leaky ReLU | Fatty | 7054 | 962 | 88.0 | 0.759 | |
Dense | 1027 | 7529 | ||||
ResNet-18 | ReLU | Fatty | 7289 | 727 | 91.9 | 0.839 |
Dense | 599 | 7957 | ||||
Sigmoid | Fatty | 7102 | 914 | 88.6 | 0.771 | |
Dense | 976 | 7580 | ||||
Tanh | Fatty | 7014 | 1002 | 87.5 | 0.749 | |
Dense | 1070 | 7486 | ||||
Leaky ReLU | Fatty | 7078 | 938 | 88.3 | 0.766 | |
Dense | 998 | 7558 |
Results Analysis
Misclassification Analysis
Experiment no | Model | Dataset | No. of samples | Miss_F | Miss_D | Mis_Acc |
---|---|---|---|---|---|---|
Experiment no. 1 | AlexNet | MIAS | 3220 | 152 | 176 | 10.2 |
Experiment no. 2 | ResNet-18 | MIAS | 1932 | 55 | 113 | 8.7 |
Experiment no. 3 | AlexNet | DDSM | 24400 | 1188 | 1206 | 9.8 |
Experiment no. 4 | ResNet-18 | DDSM | 14640 | 568 | 559 | 7.7 |
Experiment no. 5 | AlexNet | MIAS + DDSM | 27620 | 1263 | 1969 | 11.7 |
Experiment no. 6 | ResNet-18 | MIAS + DDSM | 16572 | 727 | 599 | 8.1 |
Comparative Analysis
DL model | Number of samples | Accuracy (%) | Kappa coefficient | |
---|---|---|---|---|
Proposed work | ResNet-18 | 48800 | 92.3 | 0.846 |
Valencia-Hernandez et al. [26] | Takagi–Sugeno | 1010 | 84.2 | – |
Dontchos et al. [9] | CNN | 2174 | 90.7 | – |
Clancy et al. [10] | AlexNet | 22000 | 77.0 | – |
Gandomkar et al. [16] | Inception-V3 | 3813 | 83.3 | 0.775 |
Shi et al. [15] | VGG-16 | 10304 | 83.9 | – |