Skip to content

Latest commit

 

History

History
234 lines (182 loc) · 10.7 KB

Results.md

File metadata and controls

234 lines (182 loc) · 10.7 KB

Project Steps & Results

Baseline1 - ResNet50 - SPIKaMed Full Images

✅ TOP1ACC on test = ==85.15%==

Precision Recall F1-Score Support
Dyskeratotic 0.869 0.869 0.869 23
Koilocytotic 0.889 0.64 0.744 25
Metaplastic 0.764 0.928 0.839 28
Parabasal 0.923 1 0.96 12
Superficial 0.846 0.846 0.846 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.869 0.0 0.130 0.0 0.0
Koilocytotic 0.04 0.64 0.24 0.04 0.04
Metaplastic 0.036 0.036 0.928 0.0 0.0
Parabasal 0.0 0.0 0.0 1.0 0.0
Superficial 0.0 0.0 0.0 0.0 1.0

Baseline2 - ResNet50 - SPIKaMed Cropped Images

✅ TOP1ACC on test = ==95.11%==

Precision Recall F1-Score Support
Dyskeratotic 0.975 0.951 0.963 82
Koilocytotic 0.895 0.928 0.911 83
Metaplastic 0.95 0.95 0.95 80
Parabasal 1 0.988 0.994 80
Superficial 0.988 0.988 0.988 84
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.963 0.036 0.0 0.0 0.0
Koilocytotic 0.024 0.915 0.036 0.0 0.024
Metaplastic 0.0 0.037 0.963 0.0 0.0
Parabasal 0.0 0.0 0.0 1.0 0.0
Superficial 0.012 0.0 0.0 0.0 0.988

Baseline3 - ResNet50 - Masked Region of SPIKaMeD Full Images

✅ TOP1ACC on test = ==86.13%==

Precision Recall F1-Score Support
Dyskeratotic 0.815 0.957 0.88 23
Koilocytotic 0.789 0.6 0.682 25
Metaplastic 0.815 0.786 0.8 28
Parabasal 0.769 0.833 0.8 12
Superficial 0.8 0.923 0.857 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.957 0.043 0.0 0.0 0.0
Koilocytotic 0.08 0.64 0.12 0.04 0.12
Metaplastic 0.035 0.107 0.821 0.035 0.0
Parabasal 0.167 0.0 0.0 0.833 0.0
Superficial 0.0 0.0 0.0 0.0 1.0

Generalization Test - Baseline w cropped SPIKaMed model on Smear dataset

(Train:test=1:9, 50 epochs training)

✅ TOP1ACC on test = ==57.61%==

Precision Recall F1-Score Support
Carcinoma 0.532 0.7 0.604 120
Light Dysplastic 0.680 0.685 0.683 146
Moderate Dysplastic 0.419 0.402 0.410 117
Normal Columnar 0.521 0.481 0.5 79
Normal Intermediate 0.742 0.875 0.803 56
Normal Superficiel 0.959 0.783 0.862 60
Severe Dysplastic 0.405 0.335 0.367 158
predict\truth Car- L-Dys- M-Dys- N-Col- N-Inter- N-Sup- S-Dys
Carcinoma 0.683 0.016 0.041 0.025 0.0 0.0 0.0
Light Dysplastic 0.034 0.630 0.144 0.021 0.027 0.0 0.144
Moderate Dysplastic 0.094 0.256 0.427 0.008 0.017 0.0 0.197
Normal Columnar 0.278 0.051 0.025 0.494 0.0 0.0 0.152
Normal Intermediate 0.018 0.036 0.036 0.036 0.875 0.036 0.0
Normal Superficiel 0.0 0.0 0.0 0.0 0.217 0.783 0.0
Severe Dysplastic 0.196 0.057 0.234 0.171 0.0 0.0 0.341

Model 1 - Residual Attention Network - SPIKaMeD Full Images

✅ TOP1ACC on test = ==84.16%==

Precision Recall F1-Score Support
Dyskeratotic 0.833 0.869 0.851 23
Koilocytotic 0.857 0.72 0.783 25
Metaplastic 0.862 0.893 0.877 28
Parabasal 0.846 0.917 0.88 12
Superficial 0.786 0.846 0.815 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.869 0.086 0.043 0.0 0.0
Koilocytotic 0.08 0.72 0.08 0.0 0.12
Metaplastic 0.036 0.0 0.893 0.071 0.0
Parabasal 0.083 0.0 0.0 0.917 0.0
Superficial 0.0 0.077 0.077 0.0 0.846

Model 2 - DenseNet - SPIKaMeD Full Images

✅ TOP1ACC on test = ==89.11%==

Precision Recall F1-Score Support
Dyskeratotic 0.846 0.957 0.898 23
Koilocytotic 0.864 0.76 0.809 25
Metaplastic 0.929 0.929 0.929 28
Parabasal 1 0.917 0.957 12
Superficial 0.857 0.923 0.889 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.957 0.043 0.0 0.0 0.0
Koilocytotic 0.08 0.76 0.08 0.0 0.08
Metaplastic 0.035 0.035 0.929 0.0 0.0
Parabasal 0.083 0.0 0.0 0.917 0.0
Superficial 0.0 0.077 0.0 0.0 0.923

Model 3 - DenseNet - SPIKaMeD Cropped Images

✅ TOP1ACC on test = ==95.84%==

Precision Recall F1-Score Support
Dyskeratotic 0.952 0.963 0.958 82
Koilocytotic 0.913 0.879 0.896 83
Metaplastic 0.949 0.938 0.943 80
Parabasal 1.0 0.988 0.994 80
Superficial 0.955 1.0 0.977 84
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.963 0.037 0.0 0.0 0.0
Koilocytotic 0.048 0.879 0.048 0.0 0.024
Metaplastic 0.0 0.05 0.938 0.0 0.012
Parabasal 0.0 0.0 0.0 0.988 0.012
Superficial 0.0 0.0 0.0 0.0 1.0

Model 4 - DenseNet - Masked Region of SPIKaMeD Full Images

✅ TOP1ACC on test = ==90.10%==

Precision Recall F1-Score Support
Dyskeratotic 0.913 0.913 0.913 23
Koilocytotic 0.875 0.84 0.857 25
Metaplastic 0.893 0.893 0.893 28
Parabasal 1.0 0.75 0.857 12
Superficial 0.765 1.0 0.867 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.913 0.087 0.0 0.0 0.0
Koilocytotic 0.04 0.84 0.04 0.0 0.08
Metaplastic 0.036 0.0 0.893 0.0 0.071
Parabasal 0.0 0.083 0.167 0.75 0.0
Superficial 0.0 0.0 0.0 0.0 1.0

Model 5 - Channel Attention DenseNet - SPIKaMeD Full Images

✅ TOP1ACC on test = ==91.09%==

Precision Recall F1-Score Support
Dyskeratotic 0.958 1.0 0.978 23
Koilocytotic 0.952 0.8 0.869 25
Metaplastic 0.867 0.929 0.896 28
Parabasal 1.0 1.0 1.0 12
Superficial 0.857 0.923 0.889 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 1.0 0.0 0.0 0.0 0.0
Koilocytotic 0.0 0.8 0.12 0.0 0.08
Metaplastic 0.036 0.036 0.929 0.0 0.0
Parabasal 0.0 0.0 0.0 1.0 0.0
Superficial 0.0 0.0 0.077 0.0 0.923

Model 6 - Channel Attention DenseNet - SPIKaMeD Cropped Images

✅ TOP1ACC on test = ==96.33%==

Precision Recall F1-Score Support
Dyskeratotic 0.941 0.976 0.958 82
Koilocytotic 0.907 0.939 0.923 83
Metaplastic 0.962 0.95 0.956 80
Parabasal 1.0 0.9875 0.994 80
Superficial 1.0 0.952 0.976 84
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.976 0.024 0.0 0.0 0.0
Koilocytotic 0.036 0.940 0.024 0.0 0.0
Metaplastic 0.0 0.05 0.95 0.0 0.0
Parabasal 0.0125 0.0 0.0 0.9875 0.0
Superficial 0.012 0.024 0.012 0.0 0.952

Model 7 - Channel Attention DenseNet - SPIKaMeD Masked Images

✅ TOP1ACC on test = ==87.13%==

Precision Recall F1-Score Support
Dyskeratotic 1.0 0.913 0.955 23
Koilocytotic 0.710 0.88 0.786 25
Metaplastic 0.8 0.857 0.828 28
Parabasal 1.0 0.833 0.909 12
Superficial 0.889 0.615 0.727 13
predict\truth Dys- Koi- Met- Par- Sup-
Dyskeratotic 0.913 0.043 0.043 0.0 0.0
Koilocytotic 0.0 0.88 0.08 0.0 0.04
Metaplastic 0.0 0.143 0.857 0.0 0.0
Parabasal 0.0 0.083 0.083 0.833 0.0
Superficial 0.0 0.231 0.154 0.0 0.615