Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification Neslihan Bayramoglu 1 , Juho Kannala 2 , Janne Heikkilä 1 1 Center for Machine Vision and Signal Analysis, University of Oulu, Finland. 2 Department of Computer Science Aalto University, Finland
Breast Cancer ● Most common cancer among women ● For a definitive diagnosis ● Biopsy ● Microscopic analysis Image: http://www.rafautama.com Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Sample Preparation Image: http://amida13.isi.uu.nl Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Visual Image Analysis ● Pathologists: ● pan ● focus ● zoom ● scan can entire image at high magnifications. ● Timely, costly, and subjective process! Image: www.biomatrix-intl.com Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Automated Image Analysis Methods Computer vision and machine learning methods could automate some of the tasks in the diagnostic pathology workflow ● Reduce observer variability ● Increase objectivity ● Fast and precise quantification ● Enhance the healthcare quality Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Challenges Appearance variability of hematoxylin and eosin stained sections ● variability among people ● differences in protocols between labs ● specimen orientation ● human skills in tissue preparation ● microscopy maintenance ● color variation due to differences in staining procedures Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Magnification A malignant breast tumor acquired from a single slide seen in different magnification factors: 40×, 100×, 200×, and 400× Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Magnification Most of the previous studies: ● Utilize same magnification level Classifier Model Train data Fixed magnification Test image same Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Magnification Most of the previous studies: ● Utilize same magnification level Others: ● Utilize multiple magnifications ● Different classifier for each magnification level Classifier 1 Model 1 Classifier 2 Model 2 Train data Train data magnification I magnification II Test image Test image s a s a m e m e Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Magnification Most of the previous studies: ● Utilize same magnification level Others: ● Utilize multiple magnifications ● Different classifier for each magnification level Practical limitations ● Multiple training stages ● Test time: magnification factor should be known ● Difficult to adapt for test images acquired at new magnification levels. Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Proposal Proposal 1. Classify images independent of their magnifications. 2. Multi-task classification Simultaneous recognition of magnification level and tumor class. Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Dataset Breast Cancer Histopathological Database (BreakHis) Magnification Benign Malignant Total 40X 652 1370 1995 100X 644 1437 2081 200X 623 1390 2013 400X 588 1232 1820 Total of Images 2480 5429 7909 http://web.inf.ufpr.br/vri/breast-cancer-database Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Overview Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Data augmentation Augmentation ● Rotations ● 90°, 180°, and 270°. ● Flip Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Magnification independent classification (single-task CNN) Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Multi-task classification Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Multi-task classification Multi-task loss Each output layer computes a discrete probability distribution by a softmax over the outputs of a fully connected layer. magnification L C = w bening/malignant L o s s bening/malignant + w o s s magnification ● Different weights might improve the results. ● Difficult to determine theoretically. ● Needs to be estimated empirically. Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Experiments and Results Performance Evaluation Performance Evaluation N P =Number of images of patient P N rec = Number of correctly classified images Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Experiments and Results Recognition Rate (based on patient score) (%) Method/ Magnification 40x 100x 200x 400x Average Hand crafted Features CLBP 77.4±3.8 76.4±4.5 70.2±3.6 72.8±4.9 74.2 GLCM 74.7±1.0 78.6±2.6 83.4±3.3 81.7±3.3 79.6 LBP 75.6±2.4 73.2±3.5 72.9±2.3 73.1±5.7 73.7 LPQ 73.8±5.0 72.8±5.0 74.3±6.3 73.7±5.7 73.65 ORB 74.4±1.7 69.4±0.4 69.6±3.0 67.6±1.2 70.25 PFTAS 83.8±2.0 82.1±4.9 85.1±3.1 82.3±3.8 83.33 SVM QDA Multi-task CNN 81.87±3.06 83.39±5.17 82.56±3.49 80.69±4.23 82.13 Single-CNN 83.08±2.08 83.17±3.51 84.63±2.72 82.10±4.42 83.25 Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Experiments and Results Recognition Rate (based on patient score) (%) Method/ Magnification 40x 100x 200x 400x Average Hand crafted Features CLBP 77.4±3.8 76.4±4.5 70.2±3.6 72.8±4.9 74.2 GLCM 74.7±1.0 78.6±2.6 83.4±3.3 81.7±3.3 79.6 LBP 75.6±2.4 73.2±3.5 72.9±2.3 73.1±5.7 73.7 LPQ 73.8±5.0 72.8±5.0 74.3±6.3 73.7±5.7 73.65 ORB 74.4±1.7 69.4±0.4 69.6±3.0 67.6±1.2 70.25 PFTAS 83.8±2.0 82.1±4.9 85.1±3.1 82.3±3.8 83.33 Train 40x Train 100x Train 200x Train 400x Multi-task CNN 81.87±3.06 83.39±5.17 82.56±3.49 80.69±4.23 82.13 Single-CNN 83.08±2.08 83.17±3.51 84.63±2.72 82.10±4.42 83.25 Single Train Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Summary ● Independent from microscopy magnification and faster than previous methods. ● Models are scalable . ● Multi-task CNN architecture to predict both the image magnification level and its benign/malignancy property simultaneously. ● Combine image data from many more resolution levels than four discrete magnification levels. ● Magnification level prediction could be formulated as a regression problem ● Multi-task prediction requires essentially no additional computation over single-task prediction Deep Learning for Magnification Independent Breast Cancer Histopathology Image Classification
Recommend
More recommend
