CONVOLUTIONAL AUTOENCODERS BASED FEATURE EXTRACTION FOR THYROID NODULE

CONVOLUTIONAL AUTO-ENCODERS BASED FEATURE EXTRACTION FOR THYROID NODULE ULTRASOUND IMAGE Presented by: Professor Theodore Trafalis Department of Industrial System Engineering University of Oklahoma Norman, OK, USA 17 th International Symposium on Mathematical and Computational Biology

Outline • Thyroid Nodule Images • Feature Extraction Framework • Convolutional Auto-encoders • Local Binary Patterns • Histogram of Oriented Gradients • Average Pooling • Experiments • Conclusion

Thyroid Nodule • 68% thyroid nodules are benign • Most common type of cancer: papillary cancer • Thyroid Ultrasound Reporting and Data System(TIRADS)

Deep Feature Extraction Framework • Convolutional Auto-encoders • Local Binary Pattern(LBP) • Histogram of Oriented Gradients(HOG) • Global Average Pooling

Convolutional Autoencoders (CAE)

Convolutional Auto-Encoders (CAE) Layer Width Height # filter Kernel after size/p Strides layer ool size Input 224 1 NA 1 Conv 1 224 64 (3, 3) 1 Pool 1 112 64 (2, 2) 1 Conv 2 112 32 (3, 3) 1 Pool 2 56 56 32 (2, 2) 1 Conv 3 56 56 16 (3, 3) 1 Conv 4 28 28 8 (3, 3) 1 Pool 3 28 28 8 (2, 2) 1 Encode 28 28 8 NA 1 d

Encoded Thyroid Nodule Image

Local Binary Patterns (LBP) • Visual Descriptor in Computer Vision • Texture Description Image Courtesy: Bikramjot Singh Hanzra

LBP on Ultrasound Image

Histogram of Oriented Gradients (HOG) • Visual Descriptor in Computer Vision • Shape descriptor • Directional Patterns Image Courtesy: skimage

HOG on Ultrasound Image

Average Pooling

Experiments • Data collected at East River Medical Imaging, New York, USA • Equipment: General Electric Logiq L 9 and E 9 ultrasound machines • 427 cases with 366 Benign • Assuming every single image is independent • Label Benign as “ 0” ---Bethesda label 2 • Label Malignant as “ 1”—Bethasda label 3 -6

Experiments • Stratified Cross Validation • Models tested: • • Random Forest Logistic Regression Linear Support Vector Machines RBF Support Vector Machines • Metrics: • • • Accuracy: (TN+TP)/(TN+TP+FP+FN) Negative Precision/Precision 0/Negative Predictive Value Negative Recall/Recall 0/ Specificity Positive Precision/Precision 1/ Positive Predictive Value/ Positive Recall/Recall 1/Sensitivity

Experiments Precision Classifier Accuracy 0 Random Forest Logistic Regressio n Linear SVM RBF SVM Precisio n 1 Recall 0 Recall 1 Performan ce score 0. 89 0. 97 0. 99 0. 25 2. 03 0. 66 0. 90 0. 22 0. 68 0. 54 2. 10 0. 75 0. 89 0. 25 0. 81 0. 39 2. 02 0. 64 0. 93 0. 24 0. 62 0. 72 2. 28 • Performance Score = Accuracy+Precision 0+Recall 1

Conclusion • • • CAE could extract useful features from thyroid images LBP + HOG could describe texture information of the thyroid nodule RBF SVM achieves the highest performance score and true negative rate Random Forest achieves the highest true positive rate This model might be useful for clinical practice to reduce unnecessary benign biopsies

Q&A