Building a simple diabetic retinopathy classifier with attention

Dr. Rahul Remanan

CEO, Moad Computer

Described in this article is a simple deep convolutional network (CNN), with a state-of-the-art EfficientNet backbone and an attention attention mechanism, to classify the severity of diabetic retinopathy in retinal images. The model can be trained using the publicly available diabetic retinopathy dataset on Kaggle.

Open in Kaggle

Configuration

Open in Kaggle

Imports

Open in Kaggle

​Manage training data

Specify paths for the training files

Helper function to run Linux commands within Python

Unzip the training data

Create and process training dataframe

	image	level	PatientId	path	exists	eye	level_cat
20803	26226_right	0	26226	/tmp/train/train/26226_right.jpeg	True	0	[1.0, 0.0, 0.0, 0.0, 0.0]
16437	20592_right	0	20592	/tmp/train/train/20592_right.jpeg	True	0	[1.0, 0.0, 0.0, 0.0, 0.0]
418	483_left	0	483	/tmp/train/train/483_left.jpeg	True	1	[1.0, 0.0, 0.0, 0.0, 0.0]

Histogram summary of the training dataframe

Open in Kaggle

Train-validation split

Open in Kaggle

Balance the training data

Open in Kaggle

Train and validation data generators

Open and process images

Open in Kaggle

Data generators

Open in Kaggle

Visualize the data generator outputs

Visualize train data generator

Visualize validation data generator

Open in Kaggle

Attention mechanism

Open in Kaggle

Diabetic retinopathy classifier

Open in Kaggle

Model summary

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_2 (InputLayer)            [(None, 512, 512, 3) 0                                            
__________________________________________________________________________________________________
efficientnetb0 (Functional)     (None, 16, 16, 1280) 4049571     input_2[0][0]                    
__________________________________________________________________________________________________
retina_model_bn_inp (BatchNorma (None, 16, 16, 1280) 5120        efficientnetb0[0][0]             
__________________________________________________________________________________________________
retina_model_attn_dropout_in (D (None, 16, 16, 1280) 0           retina_model_bn_inp[0][0]        
__________________________________________________________________________________________________
retina_model_attn_conv2D_64 (Co (None, 16, 16, 64)   81984       retina_model_attn_dropout_in[0][0
__________________________________________________________________________________________________
retina_model_attn_conv2D_16 (Co (None, 16, 16, 16)   1040        retina_model_attn_conv2D_64[0][0]
__________________________________________________________________________________________________
retina_model_attn_conv2D_8 (Con (None, 16, 16, 8)    136         retina_model_attn_conv2D_16[0][0]
__________________________________________________________________________________________________
retina_model_attn_conv2D_1 (Con (None, 16, 16, 1)    9           retina_model_attn_conv2D_8[0][0] 
__________________________________________________________________________________________________
retina_model_attn_final (Conv2D (None, 16, 16, 1280) 1280        retina_model_attn_conv2D_1[0][0] 
__________________________________________________________________________________________________
retina_model_attn_out (Multiply (None, 16, 16, 1280) 0           retina_model_attn_final[0][0]    
                                                                 retina_model_attn_dropout_in[0][0
__________________________________________________________________________________________________
retina_model_gap_feat (GlobalAv (None, 1280)         0           retina_model_attn_final[0][0]    
__________________________________________________________________________________________________
retina_model_gap_mask (GlobalAv (None, 1280)         0           retina_model_attn_out[0][0]      
__________________________________________________________________________________________________
retina_model_reshape_feat (Resh (None, 1, 1280)      0           retina_model_gap_feat[0][0]      
__________________________________________________________________________________________________
retina_model_reshape_mask (Resh (None, 1, 1280)      0           retina_model_gap_mask[0][0]      
__________________________________________________________________________________________________
retina_model_gap_rescale (Lambd (None, 1, 1280)      0           retina_model_reshape_feat[0][0]  
                                                                 retina_model_reshape_mask[0][0]  
__________________________________________________________________________________________________
retina_model_gap_dropout (Dropo (None, 1, 1280)      0           retina_model_gap_rescale[0][0]   
__________________________________________________________________________________________________
retina_model_fc_dense_128 (Dens (None, 1, 128)       163968      retina_model_gap_dropout[0][0]   
__________________________________________________________________________________________________
retina_model_fc_dropout (Dropou (None, 1, 128)       0           retina_model_fc_dense_128[0][0]  
__________________________________________________________________________________________________
retina_model_output (Dense)     (None, 1, 5)         645         retina_model_fc_dropout[0][0]    
==================================================================================================
Total params: 4,303,753
Trainable params: 250,342
Non-trainable params: 4,053,411
__________________________________________________________________________________________________
None

Open in Kaggle

Train model

Open in Kaggle

Visualize attention mechanism

Open in Kaggle

References:
1. This code article is forked and modified from the diabetic retinopathy detection notebook authored by @manifoldix
2. Kaggle diabetic retinopathy dataset