Transfer Learning for Cell Nuclei Classification in Histopathology - PowerPoint PPT Presentation

Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu, Janne Heikkilä Center for Machine Vision and Signal Analysis, University of Oulu, Finland.

Biomedical Image Analysis ● Histopathological image assessment ● microscopic examination of tissue ● High demand to obtain fast and precise quantification automatically. Adenoid cystic Ductal carcinoma of Ovarian cancer carcinoma the breast Stained with Hematoxylin & Eosin (H&E) Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Biomedical Image Analysis ● Automated techniques beneficial to ● find clinical assessment clues, ● produce correct diagnoses, ● reduce observer variability, ● increase objectivity. ● Deep learning could be the key method to obtain clinical acceptance Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Deep Learning for Biomedical Image Analysis ● Bottleneck: Limited amount of training data. ● Question: Could it be possible to use transfer learning and fine-tuning in biomedical image analysis to reduce the effort of manual data labeling and still obtain a full deep representation for the target task? Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Deep Learning for Biomedical Image Analysis ● Significant differences in image statistics between biomedical images and natural images ● We evaluate whether the features learned from deep CNNs trained on generic recognition tasks could generalize to biomedical tasks. Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Deep Learning for Biomedical Image Analysis ● Significant differences in image statistics between biomedical images and natural images ● We evaluate whether the features learned from deep CNNs trained on generic recognition tasks could generalize to biomedical tasks. Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Cell Nuclei Classification in Histopathology Images Data Set ● H&E stained histopathology images of colorectal adenocarcinoma. ● 20,405 manually labelled cell nuclei (training: 17,004, testing : 3401). ● Categories: Epithelial (7,772), inflammatory (6,971), fibroblast (5,712), and miscellaneous (excluded). ● Publicy available. Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Experiments VGG-16 GoogLeNet GenderNet AlexNet 13 Layers 22 Layers 3 Layers 5 Layers ● Full training: network parameters are initialized randomly. ● Raw images, no data augmentation. Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu

Results GenderNet VGG_16 0.9 0.9 0.8 0.8 Classification Accuracy Classification Accuracy 0.7 0.7 0.9 0.9 GoogLeNet AlexNet 0.8 0.8 0.7 0.7 0.6 0.6 200 400 600 800 1,000 1,200 1,400 1,600 0.5 Number of Iterations 0 200 400 600 800 1,000 1,200 1,400 1,600 0 500 1,000 1,500 2,000 2,500 3,000 Number of Iterations Number of Iterations Transfer Learning Full Training Initializing Transfer learning the network parameters Feature transferability is Deeper and fine-tuning with transferred affected by the depth of architectures provides much features can improve the network, source task, trained on bigger better results the classification and the diversity of the datasets converge than learning performance for any source data. faster. from scratch. model. Transfer Learning for Cell Nuclei Classification in Histopathology Images Neslihan Bayramoglu