Close
  Indian J Med Microbiol
 

Figure 1: (a) Dataset A included 1000 images acquired through ten slides (five benign colon and five invasive colon carcinomas). (b) From the above 1000 images from Dataset A, seven distinct image training set categories (1000 images, 500 images, 200 images, 100 images, 50 images, 30 images, and 10 images) were constructed to assess the significance of the number of training images on their respective model's accuracy. (c) A transfer-learning approach was employed to retrain the three distinct well established convolutional neural networks noted above in building models that could distinguish colonic carcinoma from normal colonic tissue. (d) The model's accuracy was then assessed through two distinct data sets “Internal Validation” is based on Dataset A's 20% of the images that were kept outside of the training phase and used for the first validation accuracy measure while the “External Validation” test set is based on Dataset B which was completely unknown to our trained images (taken from a variety of public domain sources) and used to assess each model's generalizability. (e) The performance parameters of the individual models were then compared, contrasted, and statistically analyzed

Figure 1: (a) Dataset A included 1000 images acquired through ten slides (five benign colon and five invasive colon carcinomas). (b) From the above 1000 images from Dataset A, seven distinct image training set categories (1000 images, 500 images, 200 images, 100 images, 50 images, 30 images, and 10 images) were constructed to assess the significance of the number of training images on their respective model's accuracy. (c) A transfer-learning approach was employed to retrain the three distinct well established convolutional neural networks noted above in building models that could distinguish colonic carcinoma from normal colonic tissue. (d) The model's accuracy was then assessed through two distinct data sets “Internal Validation” is based on Dataset A's 20% of the images that were kept outside of the training phase and used for the  first validation accuracy measure while the “External Validation” test set is based on Dataset B which was completely unknown to our trained images (taken from a variety of public domain sources) and used to assess each model's generalizability. (e) The performance parameters of the individual models were then compared, contrasted, and statistically analyzed