Detecting Diabetic Retinopathy With Deep Learning

Objective

Diabetic retinopathy is the leading cause of blindness in the working-age population of the developed world. The condition is estimated to affect over 93 million people.

The need for a comprehensive and automated method of diabetic retinopathy screening has long been recognized, and previous efforts have made good progress using image classification, pattern recognition, and machine learning. With photos of eyes as input, the goal of this capstone is to create a new model, ideally resulting in realistic clinical potential.

The motivations for this project are twofold:

• Image classification has been a personal interest for years, in addition to classification on a large scale data set.

• Time is lost between patients getting their eyes scanned (shown below), having their images analyzed by doctors, and scheduling a follow-up appointment. By processing images in real-time, EyeNet would allow people to seek & schedule treatment the same day.

Table of Contents

1. Data
2. Extraction and Preprocessing
   • Download Images to Google Colab
   • Resize Images
   • Checking Blurness of Images
   • Data Augmentation
3. CNN Architecture
4. Results
5. References
6. Authors

Data

The data originates from a 2015 Kaggle competition. However, is an atypical Kaggle dataset. In most Kaggle competitions, the data has already been cleaned, giving the data scientist very little to preprocess. With this dataset, this isn't the case.

All images are taken of different people, using different cameras, and of different sizes. Pertaining to the preprocessing section, this data is extremely noisy, and requires multiple preprocessing steps to get all images to a useable format for training a model.

The training data is comprised of 35,126 images, which are augmented during preprocessing.

Prerequisites

You'll need to install:

• Anaconda
• Python (Minimum 3)
• pandas
• Seaborn
• Fast.ai

Extraction and Preprocessing

The preprocessing pipeline is the following:

Download Images to Google Colab

Google Colab was used as a platform for this dataset.

Resize All Images

The given images were mostly very large. ~2000 by ~2000. So to keep things unifrom all images were scaled down to 512 by 512.

Checking Blurness of Images

The method is simple. Straightforward. Has sound reasoning. And can be implemented in only a single line of code:

cv2.Laplacian(image, cv2.CV_64F).var()

We simply take a single channel of an image and convolve it with the following 3 x 3 kernel:

0	1	0
1	-4	1
0	1	0

And then take the variance (i.e. standard deviation squared) of the response. If the variance falls below a pre-defined threshold, then the image is considered blurry; otherwise, the image is not blurry. Here is the paper with talks about it's implementation, Ariation of the Laplacian by Pech-Pacheco et al. in their 2000 ICPR paper, Diatom autofocusing in brightfield microscopy: a comparative study.

Data Augmentation

All images were rotated and mirrored.Images without retinopathy were mirrored; images that had retinopathy were mirrored, and rotated 90, 120, 180, and 270 degrees.

The first images show two pairs of eyes, along with the black borders. Notice in the cropping and rotations how the majority of noise is removed.

Neural Network Architecture

Layer (type)	Output Shape	Param #
Conv2d-1	[-1, 64, 256, 256]	9,408
BatchNorm2d-2	[-1, 64, 256, 256]	128
ReLU-3	[-1, 64, 256, 256]	0
MaxPool2d-4	[-1, 64, 128, 128]	0
BatchNorm2d-5	[-1, 64, 128, 128]	128
ReLU-6	[-1, 64, 128, 128]	0
Conv2d-7	[-1, 128, 128, 128]	8192
BatchNorm2d-8	[-1, 128, 128, 128]	256
ReLU-9	[-1, 128, 128, 128]	0
Conv2d-10	[-1, 32, 128, 128]	36864
BatchNorm2d-11	[-1, 96, 128, 128]	192
ReLU-12	[-1, 96, 128, 128]	0
Conv2d-13	[-1, 128, 128, 128]	12288
BatchNorm2d-14	[-1, 128, 128, 128]	256
ReLU-15	[-1, 128, 128, 128]	0
Conv2d-16	[-1, 32, 128, 128]	36864
BatchNorm2d-17	[-1, 128, 128, 128]	256
ReLU-18	[-1, 128, 128, 128]	0
Conv2d-19	[-1, 128, 128, 128]	16384
BatchNorm2d-20	[-1, 128, 128, 128]	256
ReLU-21	[-1, 128, 128, 128]	0
Conv2d-22	[-1, 32, 128, 128]	36864
BatchNorm2d-23	[-1, 160, 128, 128]	320
ReLU-24	[-1, 160, 128, 128]	0
Conv2d-25	[-1, 128, 128, 128]	20480
BatchNorm2d-26	[-1, 128, 128, 128]	256
ReLU-27	[-1, 128, 128, 128]	0
Conv2d-28	[-1, 32, 128, 128]	36864
BatchNorm2d-29	[-1, 192, 128, 128]	384
ReLU-30	[-1, 192, 128, 128]	0
Conv2d-31	[-1, 128, 128, 128]	24576
BatchNorm2d-32	[-1, 128, 128, 128]	256
ReLU-33	[-1, 128, 128, 128]	0
Conv2d-34	[-1, 32, 128, 128]	36864
BatchNorm2d-35	[-1, 224, 128, 128]	448
ReLU-36	[-1, 224, 128, 128]	0
Conv2d-37	[-1, 128, 128, 128]	28672
BatchNorm2d-38	[-1, 128, 128, 128]	256
ReLU-39	[-1, 128, 128, 128]	0
Conv2d-40	[-1, 32, 128, 128]	36864
BatchNorm2d-41	[-1, 256, 128, 128]	512
ReLU-42	[-1, 256, 128, 128]	0
Conv2d-43	[-1, 128, 128, 128]	32768
AvgPool2d-44	[-1, 128, 64, 64]	0
BatchNorm2d-45	[-1, 128, 64, 64]	256
ReLU-46	[-1, 128, 64, 64]	0
Conv2d-47	[-1, 128, 64, 64]	16384
BatchNorm2d-48	[-1, 128, 64, 64]	256
ReLU-49	[-1, 128, 64, 64]	0
Conv2d-50	[-1, 32, 64, 64]	36864
BatchNorm2d-51	[-1, 160, 64, 64]	320
ReLU-52	[-1, 160, 64, 64]	0
Conv2d-53	[-1, 128, 64, 64]	20480
BatchNorm2d-54	[-1, 128, 64, 64]	256
ReLU-55	[-1, 128, 64, 64]	0
Conv2d-56	[-1, 32, 64, 64]	36864
BatchNorm2d-57	[-1, 192, 64, 64]	384
ReLU-58	[-1, 192, 64, 64]	0
Conv2d-59	[-1, 128, 64, 64]	24576
BatchNorm2d-60	[-1, 128, 64, 64]	256
ReLU-61	[-1, 128, 64, 64]	0
Conv2d-62	[-1, 32, 64, 64]	36864
BatchNorm2d-63	[-1, 224, 64, 64]	448
ReLU-64	[-1, 224, 64, 64]	0
Conv2d-65	[-1, 128, 64, 64]	28672
BatchNorm2d-66	[-1, 128, 64, 64]	256
ReLU-67	[-1, 128, 64, 64]	0
Conv2d-68	[-1, 32, 64, 64]	36864
BatchNorm2d-69	[-1, 256, 64, 64]	512
ReLU-70	[-1, 256, 64, 64]	0
Conv2d-71	[-1, 128, 64, 64]	32768
BatchNorm2d-72	[-1, 128, 64, 64]	256
ReLU-73	[-1, 128, 64, 64]	0
Conv2d-74	[-1, 32, 64, 64]	36864
BatchNorm2d-75	[-1, 288, 64, 64]	576
ReLU-76	[-1, 288, 64, 64]	0
Conv2d-77	[-1, 128, 64, 64]	36864
BatchNorm2d-78	[-1, 128, 64, 64]	256
ReLU-79	[-1, 128, 64, 64]	0
Conv2d-80	[-1, 32, 64, 64]	36864
BatchNorm2d-81	[-1, 320, 64, 64]	640
ReLU-82	[-1, 320, 64, 64]	0
Conv2d-83	[-1, 128, 64, 64]	40960
BatchNorm2d-84	[-1, 128, 64, 64]	256
ReLU-85	[-1, 128, 64, 64]	0
Conv2d-86	[-1, 32, 64, 64]	36864
BatchNorm2d-87	[-1, 352, 64, 64]	704
ReLU-88	[-1, 352, 64, 64]	0
Conv2d-89	[-1, 128, 64, 64]	45056
BatchNorm2d-90	[-1, 128, 64, 64]	256
ReLU-91	[-1, 128, 64, 64]	0
Conv2d-92	[-1, 32, 64, 64]	36864
BatchNorm2d-93	[-1, 384, 64, 64]	768
ReLU-94	[-1, 384, 64, 64]	0
Conv2d-95	[-1, 128, 64, 64]	49152
BatchNorm2d-96	[-1, 128, 64, 64]	256
ReLU-97	[-1, 128, 64, 64]	0
Conv2d-98	[-1, 32, 64, 64]	36864
BatchNorm2d-99	[-1, 416, 64, 64]	832
ReLU-100	[-1, 416, 64, 64]	0
Conv2d-101	[-1, 128, 64, 64]	53248
BatchNorm2d-102	[-1, 128, 64, 64]	256
ReLU-103	[-1, 128, 64, 64]	0
Conv2d-104	[-1, 32, 64, 64]	36864
BatchNorm2d-105	[-1, 448, 64, 64]	896
ReLU-106	[-1, 448, 64, 64]	0
Conv2d-107	[-1, 128, 64, 64]	57344
BatchNorm2d-108	[-1, 128, 64, 64]	256
ReLU-109	[-1, 128, 64, 64]	0
Conv2d-110	[-1, 32, 64, 64]	36864
BatchNorm2d-111	[-1, 480, 64, 64]	960
ReLU-112	[-1, 480, 64, 64]	0
Conv2d-113	[-1, 128, 64, 64]	61440
BatchNorm2d-114	[-1, 128, 64, 64]	256
ReLU-115	[-1, 128, 64, 64]	0
Conv2d-116	[-1, 32, 64, 64]	36864
BatchNorm2d-117	[-1, 512, 64, 64]	1024
ReLU-118	[-1, 512, 64, 64]	0
Conv2d-119	[-1, 256, 64, 64]	131072
AvgPool2d-120	[-1, 256, 32, 32]	0
BatchNorm2d-121	[-1, 256, 32, 32]	512
ReLU-122	[-1, 256, 32, 32]	0
Conv2d-123	[-1, 128, 32, 32]	32768
BatchNorm2d-124	[-1, 128, 32, 32]	256
ReLU-125	[-1, 128, 32, 32]	0
Conv2d-126	[-1, 32, 32, 32]	36864
BatchNorm2d-127	[-1, 288, 32, 32]	576
ReLU-128	[-1, 288, 32, 32]	0
Conv2d-129	[-1, 128, 32, 32]	36864
BatchNorm2d-130	[-1, 128, 32, 32]	256
ReLU-131	[-1, 128, 32, 32]	0
Conv2d-132	[-1, 32, 32, 32]	36864
BatchNorm2d-133	[-1, 320, 32, 32]	640
ReLU-134	[-1, 320, 32, 32]	0
Conv2d-135	[-1, 128, 32, 32]	40960
BatchNorm2d-136	[-1, 128, 32, 32]	256
ReLU-137	[-1, 128, 32, 32]	0
Conv2d-138	[-1, 32, 32, 32]	36864
BatchNorm2d-139	[-1, 352, 32, 32]	704
ReLU-140	[-1, 352, 32, 32]	0
Conv2d-141	[-1, 128, 32, 32]	45056
BatchNorm2d-142	[-1, 128, 32, 32]	256
ReLU-143	[-1, 128, 32, 32]	0
Conv2d-144	[-1, 32, 32, 32]	36864
BatchNorm2d-145	[-1, 384, 32, 32]	768
ReLU-146	[-1, 384, 32, 32]	0
Conv2d-147	[-1, 128, 32, 32]	49152
BatchNorm2d-148	[-1, 128, 32, 32]	256
ReLU-149	[-1, 128, 32, 32]	0
Conv2d-150	[-1, 32, 32, 32]	36864
BatchNorm2d-151	[-1, 416, 32, 32]	832
ReLU-152	[-1, 416, 32, 32]	0
Conv2d-153	[-1, 128, 32, 32]	53248
BatchNorm2d-154	[-1, 128, 32, 32]	256
ReLU-155	[-1, 128, 32, 32]	0
Conv2d-156	[-1, 32, 32, 32]	36864
BatchNorm2d-157	[-1, 448, 32, 32]	896
ReLU-158	[-1, 448, 32, 32]	0
Conv2d-159	[-1, 128, 32, 32]	57344
BatchNorm2d-160	[-1, 128, 32, 32]	256
ReLU-161	[-1, 128, 32, 32]	0
Conv2d-162	[-1, 32, 32, 32]	36864
BatchNorm2d-163	[-1, 480, 32, 32]	960
ReLU-164	[-1, 480, 32, 32]	0
Conv2d-165	[-1, 128, 32, 32]	61440
BatchNorm2d-166	[-1, 128, 32, 32]	256
ReLU-167	[-1, 128, 32, 32]	0
Conv2d-168	[-1, 32, 32, 32]	36864
BatchNorm2d-169	[-1, 512, 32, 32]	1024
ReLU-170	[-1, 512, 32, 32]	0
Conv2d-171	[-1, 128, 32, 32]	65536
BatchNorm2d-172	[-1, 128, 32, 32]	256
ReLU-173	[-1, 128, 32, 32]	0
Conv2d-174	[-1, 32, 32, 32]	36864
BatchNorm2d-175	[-1, 544, 32, 32]	1088
ReLU-176	[-1, 544, 32, 32]	0
Conv2d-177	[-1, 128, 32, 32]	69632
BatchNorm2d-178	[-1, 128, 32, 32]	256
ReLU-179	[-1, 128, 32, 32]	0
Conv2d-180	[-1, 32, 32, 32]	36864
BatchNorm2d-181	[-1, 576, 32, 32]	1152
ReLU-182	[-1, 576, 32, 32]	0
Conv2d-183	[-1, 128, 32, 32]	73728
BatchNorm2d-184	[-1, 128, 32, 32]	256
ReLU-185	[-1, 128, 32, 32]	0
Conv2d-186	[-1, 32, 32, 32]	36864
BatchNorm2d-187	[-1, 608, 32, 32]	1216
ReLU-188	[-1, 608, 32, 32]	0
Conv2d-189	[-1, 128, 32, 32]	77824
BatchNorm2d-190	[-1, 128, 32, 32]	256
ReLU-191	[-1, 128, 32, 32]	0
Conv2d-192	[-1, 32, 32, 32]	36864
BatchNorm2d-193	[-1, 640, 32, 32]	1280
ReLU-194	[-1, 640, 32, 32]	0
Conv2d-195	[-1, 128, 32, 32]	81920
BatchNorm2d-196	[-1, 128, 32, 32]	256
ReLU-197	[-1, 128, 32, 32]	0
Conv2d-198	[-1, 32, 32, 32]	36864
BatchNorm2d-199	[-1, 672, 32, 32]	1344
ReLU-200	[-1, 672, 32, 32]	0
Conv2d-201	[-1, 128, 32, 32]	86016
BatchNorm2d-202	[-1, 128, 32, 32]	256
ReLU-203	[-1, 128, 32, 32]	0
Conv2d-204	[-1, 32, 32, 32]	36864
BatchNorm2d-205	[-1, 704, 32, 32]	1408
ReLU-206	[-1, 704, 32, 32]	0
Conv2d-207	[-1, 128, 32, 32]	90112
BatchNorm2d-208	[-1, 128, 32, 32]	256
ReLU-209	[-1, 128, 32, 32]	0
Conv2d-210	[-1, 32, 32, 32]	36864
BatchNorm2d-211	[-1, 736, 32, 32]	1472
ReLU-212	[-1, 736, 32, 32]	0
Conv2d-213	[-1, 128, 32, 32]	94208
BatchNorm2d-214	[-1, 128, 32, 32]	256
ReLU-215	[-1, 128, 32, 32]	0
Conv2d-216	[-1, 32, 32, 32]	36864
BatchNorm2d-217	[-1, 768, 32, 32]	1536
ReLU-218	[-1, 768, 32, 32]	0
Conv2d-219	[-1, 128, 32, 32]	98304
BatchNorm2d-220	[-1, 128, 32, 32]	256
ReLU-221	[-1, 128, 32, 32]	0
Conv2d-222	[-1, 32, 32, 32]	36864
BatchNorm2d-223	[-1, 800, 32, 32]	1600
ReLU-224	[-1, 800, 32, 32]	0
Conv2d-225	[-1, 128, 32, 32]	102400
BatchNorm2d-226	[-1, 128, 32, 32]	256
ReLU-227	[-1, 128, 32, 32]	0
Conv2d-228	[-1, 32, 32, 32]	36864
BatchNorm2d-229	[-1, 832, 32, 32]	1664
ReLU-230	[-1, 832, 32, 32]	0
Conv2d-231	[-1, 128, 32, 32]	106496
BatchNorm2d-232	[-1, 128, 32, 32]	256
ReLU-233	[-1, 128, 32, 32]	0
Conv2d-234	[-1, 32, 32, 32]	36864
BatchNorm2d-235	[-1, 864, 32, 32]	1728
ReLU-236	[-1, 864, 32, 32]	0
Conv2d-237	[-1, 128, 32, 32]	110592
BatchNorm2d-238	[-1, 128, 32, 32]	256
ReLU-239	[-1, 128, 32, 32]	0
Conv2d-240	[-1, 32, 32, 32]	36864
BatchNorm2d-241	[-1, 896, 32, 32]	1792
ReLU-242	[-1, 896, 32, 32]	0
Conv2d-243	[-1, 128, 32, 32]	114688
BatchNorm2d-244	[-1, 128, 32, 32]	256
ReLU-245	[-1, 128, 32, 32]	0
Conv2d-246	[-1, 32, 32, 32]	36864
BatchNorm2d-247	[-1, 928, 32, 32]	1856
ReLU-248	[-1, 928, 32, 32]	0
Conv2d-249	[-1, 128, 32, 32]	118784
BatchNorm2d-250	[-1, 128, 32, 32]	256
ReLU-251	[-1, 128, 32, 32]	0
Conv2d-252	[-1, 32, 32, 32]	36864
BatchNorm2d-253	[-1, 960, 32, 32]	1920
ReLU-254	[-1, 960, 32, 32]	0
Conv2d-255	[-1, 128, 32, 32]	122880
BatchNorm2d-256	[-1, 128, 32, 32]	256
ReLU-257	[-1, 128, 32, 32]	0
Conv2d-258	[-1, 32, 32, 32]	36864
BatchNorm2d-259	[-1, 992, 32, 32]	1984
ReLU-260	[-1, 992, 32, 32]	0
Conv2d-261	[-1, 128, 32, 32]	126976
BatchNorm2d-262	[-1, 128, 32, 32]	256
ReLU-263	[-1, 128, 32, 32]	0
Conv2d-264	[-1, 32, 32, 32]	36864
BatchNorm2d-265	[-1, 1024, 32, 32]	2048
ReLU-266	[-1, 1024, 32, 32]	0
Conv2d-267	[-1, 128, 32, 32]	131072
BatchNorm2d-268	[-1, 128, 32, 32]	256
ReLU-269	[-1, 128, 32, 32]	0
Conv2d-270	[-1, 32, 32, 32]	36864
BatchNorm2d-271	[-1, 1056, 32, 32]	2112
ReLU-272	[-1, 1056, 32, 32]	0
Conv2d-273	[-1, 128, 32, 32]	135168
BatchNorm2d-274	[-1, 128, 32, 32]	256
ReLU-275	[-1, 128, 32, 32]	0
Conv2d-276	[-1, 32, 32, 32]	36864
BatchNorm2d-277	[-1, 1088, 32, 32]	2176
ReLU-278	[-1, 1088, 32, 32]	0
Conv2d-279	[-1, 128, 32, 32]	139264
BatchNorm2d-280	[-1, 128, 32, 32]	256
ReLU-281	[-1, 128, 32, 32]	0
Conv2d-282	[-1, 32, 32, 32]	36864
BatchNorm2d-283	[-1, 1120, 32, 32]	2240
ReLU-284	[-1, 1120, 32, 32]	0
Conv2d-285	[-1, 128, 32, 32]	143360
BatchNorm2d-286	[-1, 128, 32, 32]	256
ReLU-287	[-1, 128, 32, 32]	0
Conv2d-288	[-1, 32, 32, 32]	36864
BatchNorm2d-289	[-1, 1152, 32, 32]	2304
ReLU-290	[-1, 1152, 32, 32]	0
Conv2d-291	[-1, 128, 32, 32]	147456
BatchNorm2d-292	[-1, 128, 32, 32]	256
ReLU-293	[-1, 128, 32, 32]	0
Conv2d-294	[-1, 32, 32, 32]	36864
BatchNorm2d-295	[-1, 1184, 32, 32]	2368
ReLU-296	[-1, 1184, 32, 32]	0
Conv2d-297	[-1, 128, 32, 32]	151552
BatchNorm2d-298	[-1, 128, 32, 32]	256
ReLU-299	[-1, 128, 32, 32]	0
Conv2d-300	[-1, 32, 32, 32]	36864
BatchNorm2d-301	[-1, 1216, 32, 32]	2432
ReLU-302	[-1, 1216, 32, 32]	0
Conv2d-303	[-1, 128, 32, 32]	155648
BatchNorm2d-304	[-1, 128, 32, 32]	256
ReLU-305	[-1, 128, 32, 32]	0
Conv2d-306	[-1, 32, 32, 32]	36864
BatchNorm2d-307	[-1, 1248, 32, 32]	2496
ReLU-308	[-1, 1248, 32, 32]	0
Conv2d-309	[-1, 128, 32, 32]	159744
BatchNorm2d-310	[-1, 128, 32, 32]	256
ReLU-311	[-1, 128, 32, 32]	0
Conv2d-312	[-1, 32, 32, 32]	36864
BatchNorm2d-313	[-1, 1280, 32, 32]	2560
ReLU-314	[-1, 1280, 32, 32]	0
Conv2d-315	[-1, 128, 32, 32]	163840
BatchNorm2d-316	[-1, 128, 32, 32]	256
ReLU-317	[-1, 128, 32, 32]	0
Conv2d-318	[-1, 32, 32, 32]	36864
BatchNorm2d-319	[-1, 1312, 32, 32]	2624
ReLU-320	[-1, 1312, 32, 32]	0
Conv2d-321	[-1, 128, 32, 32]	167936
BatchNorm2d-322	[-1, 128, 32, 32]	256
ReLU-323	[-1, 128, 32, 32]	0
Conv2d-324	[-1, 32, 32, 32]	36864
BatchNorm2d-325	[-1, 1344, 32, 32]	2688
ReLU-326	[-1, 1344, 32, 32]	0
Conv2d-327	[-1, 128, 32, 32]	172032
BatchNorm2d-328	[-1, 128, 32, 32]	256
ReLU-329	[-1, 128, 32, 32]	0
Conv2d-330	[-1, 32, 32, 32]	36864
BatchNorm2d-331	[-1, 1376, 32, 32]	2752
ReLU-332	[-1, 1376, 32, 32]	0
Conv2d-333	[-1, 128, 32, 32]	176128
BatchNorm2d-334	[-1, 128, 32, 32]	256
ReLU-335	[-1, 128, 32, 32]	0
Conv2d-336	[-1, 32, 32, 32]	36864
BatchNorm2d-337	[-1, 1408, 32, 32]	2816
ReLU-338	[-1, 1408, 32, 32]	0
Conv2d-339	[-1, 128, 32, 32]	180224
BatchNorm2d-340	[-1, 128, 32, 32]	256
ReLU-341	[-1, 128, 32, 32]	0
Conv2d-342	[-1, 32, 32, 32]	36864
BatchNorm2d-343	[-1, 1440, 32, 32]	2880
ReLU-344	[-1, 1440, 32, 32]	0
Conv2d-345	[-1, 128, 32, 32]	184320
BatchNorm2d-346	[-1, 128, 32, 32]	256
ReLU-347	[-1, 128, 32, 32]	0
Conv2d-348	[-1, 32, 32, 32]	36864
BatchNorm2d-349	[-1, 1472, 32, 32]	2944
ReLU-350	[-1, 1472, 32, 32]	0
Conv2d-351	[-1, 128, 32, 32]	188416
BatchNorm2d-352	[-1, 128, 32, 32]	256
ReLU-353	[-1, 128, 32, 32]	0
Conv2d-354	[-1, 32, 32, 32]	36864
BatchNorm2d-355	[-1, 1504, 32, 32]	3008
ReLU-356	[-1, 1504, 32, 32]	0
Conv2d-357	[-1, 128, 32, 32]	192512
BatchNorm2d-358	[-1, 128, 32, 32]	256
ReLU-359	[-1, 128, 32, 32]	0
Conv2d-360	[-1, 32, 32, 32]	36864
BatchNorm2d-361	[-1, 1536, 32, 32]	3072
ReLU-362	[-1, 1536, 32, 32]	0
Conv2d-363	[-1, 128, 32, 32]	196608
BatchNorm2d-364	[-1, 128, 32, 32]	256
ReLU-365	[-1, 128, 32, 32]	0
Conv2d-366	[-1, 32, 32, 32]	36864
BatchNorm2d-367	[-1, 1568, 32, 32]	3136
ReLU-368	[-1, 1568, 32, 32]	0
Conv2d-369	[-1, 128, 32, 32]	200704
BatchNorm2d-370	[-1, 128, 32, 32]	256
ReLU-371	[-1, 128, 32, 32]	0
Conv2d-372	[-1, 32, 32, 32]	36864
BatchNorm2d-373	[-1, 1600, 32, 32]	3200
ReLU-374	[-1, 1600, 32, 32]	0
Conv2d-375	[-1, 128, 32, 32]	204800
BatchNorm2d-376	[-1, 128, 32, 32]	256
ReLU-377	[-1, 128, 32, 32]	0
Conv2d-378	[-1, 32, 32, 32]	36864
BatchNorm2d-379	[-1, 1632, 32, 32]	3264
ReLU-380	[-1, 1632, 32, 32]	0
Conv2d-381	[-1, 128, 32, 32]	208896
BatchNorm2d-382	[-1, 128, 32, 32]	256
ReLU-383	[-1, 128, 32, 32]	0
Conv2d-384	[-1, 32, 32, 32]	36864
BatchNorm2d-385	[-1, 1664, 32, 32]	3328
ReLU-386	[-1, 1664, 32, 32]	0
Conv2d-387	[-1, 128, 32, 32]	212992
BatchNorm2d-388	[-1, 128, 32, 32]	256
ReLU-389	[-1, 128, 32, 32]	0
Conv2d-390	[-1, 32, 32, 32]	36864
BatchNorm2d-391	[-1, 1696, 32, 32]	3392
ReLU-392	[-1, 1696, 32, 32]	0
Conv2d-393	[-1, 128, 32, 32]	217088
BatchNorm2d-394	[-1, 128, 32, 32]	256
ReLU-395	[-1, 128, 32, 32]	0
Conv2d-396	[-1, 32, 32, 32]	36864
BatchNorm2d-397	[-1, 1728, 32, 32]	3456
ReLU-398	[-1, 1728, 32, 32]	0
Conv2d-399	[-1, 128, 32, 32]	221184
BatchNorm2d-400	[-1, 128, 32, 32]	256
ReLU-401	[-1, 128, 32, 32]	0
Conv2d-402	[-1, 32, 32, 32]	36864
BatchNorm2d-403	[-1, 1760, 32, 32]	3520
ReLU-404	[-1, 1760, 32, 32]	0
Conv2d-405	[-1, 128, 32, 32]	225280
BatchNorm2d-406	[-1, 128, 32, 32]	256
ReLU-407	[-1, 128, 32, 32]	0
Conv2d-408	[-1, 32, 32, 32]	36864
BatchNorm2d-409	[-1, 1792, 32, 32]	3584
ReLU-410	[-1, 1792, 32, 32]	0
Conv2d-411	[-1, 896, 32, 32]	1605632
AvgPool2d-412	[-1, 896, 16, 16]	0
BatchNorm2d-413	[-1, 896, 16, 16]	1792
ReLU-414	[-1, 896, 16, 16]	0
Conv2d-415	[-1, 128, 16, 16]	114688
BatchNorm2d-416	[-1, 128, 16, 16]	256
ReLU-417	[-1, 128, 16, 16]	0
Conv2d-418	[-1, 32, 16, 16]	36864
BatchNorm2d-419	[-1, 928, 16, 16]	1856
ReLU-420	[-1, 928, 16, 16]	0
Conv2d-421	[-1, 128, 16, 16]	118784
BatchNorm2d-422	[-1, 128, 16, 16]	256
ReLU-423	[-1, 128, 16, 16]	0
Conv2d-424	[-1, 32, 16, 16]	36864
BatchNorm2d-425	[-1, 960, 16, 16]	1920
ReLU-426	[-1, 960, 16, 16]	0
Conv2d-427	[-1, 128, 16, 16]	122880
BatchNorm2d-428	[-1, 128, 16, 16]	256
ReLU-429	[-1, 128, 16, 16]	0
Conv2d-430	[-1, 32, 16, 16]	36864
BatchNorm2d-431	[-1, 992, 16, 16]	1984
ReLU-432	[-1, 992, 16, 16]	0
Conv2d-433	[-1, 128, 16, 16]	126976
BatchNorm2d-434	[-1, 128, 16, 16]	256
ReLU-435	[-1, 128, 16, 16]	0
Conv2d-436	[-1, 32, 16, 16]	36864
BatchNorm2d-437	[-1, 1024, 16, 16]	2048
ReLU-438	[-1, 1024, 16, 16]	0
Conv2d-439	[-1, 128, 16, 16]	131072
BatchNorm2d-440	[-1, 128, 16, 16]	256
ReLU-441	[-1, 128, 16, 16]	0
Conv2d-442	[-1, 32, 16, 16]	36864
BatchNorm2d-443	[-1, 1056, 16, 16]	2112
ReLU-444	[-1, 1056, 16, 16]	0
Conv2d-445	[-1, 128, 16, 16]	135168
BatchNorm2d-446	[-1, 128, 16, 16]	256
ReLU-447	[-1, 128, 16, 16]	0
Conv2d-448	[-1, 32, 16, 16]	36864
BatchNorm2d-449	[-1, 1088, 16, 16]	2176
ReLU-450	[-1, 1088, 16, 16]	0
Conv2d-451	[-1, 128, 16, 16]	139264
BatchNorm2d-452	[-1, 128, 16, 16]	256
ReLU-453	[-1, 128, 16, 16]	0
Conv2d-454	[-1, 32, 16, 16]	36864
BatchNorm2d-455	[-1, 1120, 16, 16]	2240
ReLU-456	[-1, 1120, 16, 16]	0
Conv2d-457	[-1, 128, 16, 16]	143360
BatchNorm2d-458	[-1, 128, 16, 16]	256
ReLU-459	[-1, 128, 16, 16]	0
Conv2d-460	[-1, 32, 16, 16]	36864
BatchNorm2d-461	[-1, 1152, 16, 16]	2304
ReLU-462	[-1, 1152, 16, 16]	0
Conv2d-463	[-1, 128, 16, 16]	147456
BatchNorm2d-464	[-1, 128, 16, 16]	256
ReLU-465	[-1, 128, 16, 16]	0
Conv2d-466	[-1, 32, 16, 16]	36864
BatchNorm2d-467	[-1, 1184, 16, 16]	2368
ReLU-468	[-1, 1184, 16, 16]	0
Conv2d-469	[-1, 128, 16, 16]	151552
BatchNorm2d-470	[-1, 128, 16, 16]	256
ReLU-471	[-1, 128, 16, 16]	0
Conv2d-472	[-1, 32, 16, 16]	36864
BatchNorm2d-473	[-1, 1216, 16, 16]	2432
ReLU-474	[-1, 1216, 16, 16]	0
Conv2d-475	[-1, 128, 16, 16]	155648
BatchNorm2d-476	[-1, 128, 16, 16]	256
ReLU-477	[-1, 128, 16, 16]	0
Conv2d-478	[-1, 32, 16, 16]	36864
BatchNorm2d-479	[-1, 1248, 16, 16]	2496
ReLU-480	[-1, 1248, 16, 16]	0
Conv2d-481	[-1, 128, 16, 16]	159744
BatchNorm2d-482	[-1, 128, 16, 16]	256
ReLU-483	[-1, 128, 16, 16]	0
Conv2d-484	[-1, 32, 16, 16]	36864
BatchNorm2d-485	[-1, 1280, 16, 16]	2560
ReLU-486	[-1, 1280, 16, 16]	0
Conv2d-487	[-1, 128, 16, 16]	163840
BatchNorm2d-488	[-1, 128, 16, 16]	256
ReLU-489	[-1, 128, 16, 16]	0
Conv2d-490	[-1, 32, 16, 16]	36864
BatchNorm2d-491	[-1, 1312, 16, 16]	2624
ReLU-492	[-1, 1312, 16, 16]	0
Conv2d-493	[-1, 128, 16, 16]	167936
BatchNorm2d-494	[-1, 128, 16, 16]	256
ReLU-495	[-1, 128, 16, 16]	0
Conv2d-496	[-1, 32, 16, 16]	36864
BatchNorm2d-497	[-1, 1344, 16, 16]	2688
ReLU-498	[-1, 1344, 16, 16]	0
Conv2d-499	[-1, 128, 16, 16]	172032
BatchNorm2d-500	[-1, 128, 16, 16]	256
ReLU-501	[-1, 128, 16, 16]	0
Conv2d-502	[-1, 32, 16, 16]	36864
BatchNorm2d-503	[-1, 1376, 16, 16]	2752
ReLU-504	[-1, 1376, 16, 16]	0
Conv2d-505	[-1, 128, 16, 16]	176128
BatchNorm2d-506	[-1, 128, 16, 16]	256
ReLU-507	[-1, 128, 16, 16]	0
Conv2d-508	[-1, 32, 16, 16]	36864
BatchNorm2d-509	[-1, 1408, 16, 16]	2816
ReLU-510	[-1, 1408, 16, 16]	0
Conv2d-511	[-1, 128, 16, 16]	180224
BatchNorm2d-512	[-1, 128, 16, 16]	256
ReLU-513	[-1, 128, 16, 16]	0
Conv2d-514	[-1, 32, 16, 16]	36864
BatchNorm2d-515	[-1, 1440, 16, 16]	2880
ReLU-516	[-1, 1440, 16, 16]	0
Conv2d-517	[-1, 128, 16, 16]	184320
BatchNorm2d-518	[-1, 128, 16, 16]	256
ReLU-519	[-1, 128, 16, 16]	0
Conv2d-520	[-1, 32, 16, 16]	36864
BatchNorm2d-521	[-1, 1472, 16, 16]	2944
ReLU-522	[-1, 1472, 16, 16]	0
Conv2d-523	[-1, 128, 16, 16]	188416
BatchNorm2d-524	[-1, 128, 16, 16]	256
ReLU-525	[-1, 128, 16, 16]	0
Conv2d-526	[-1, 32, 16, 16]	36864
BatchNorm2d-527	[-1, 1504, 16, 16]	3008
ReLU-528	[-1, 1504, 16, 16]	0
Conv2d-529	[-1, 128, 16, 16]	192512
BatchNorm2d-530	[-1, 128, 16, 16]	256
ReLU-531	[-1, 128, 16, 16]	0
Conv2d-532	[-1, 32, 16, 16]	36864
BatchNorm2d-533	[-1, 1536, 16, 16]	3072
ReLU-534	[-1, 1536, 16, 16]	0
Conv2d-535	[-1, 128, 16, 16]	196608
BatchNorm2d-536	[-1, 128, 16, 16]	256
ReLU-537	[-1, 128, 16, 16]	0
Conv2d-538	[-1, 32, 16, 16]	36864
BatchNorm2d-539	[-1, 1568, 16, 16]	3136
ReLU-540	[-1, 1568, 16, 16]	0
Conv2d-541	[-1, 128, 16, 16]	200704
BatchNorm2d-542	[-1, 128, 16, 16]	256
ReLU-543	[-1, 128, 16, 16]	0
Conv2d-544	[-1, 32, 16, 16]	36864
BatchNorm2d-545	[-1, 1600, 16, 16]	3200
ReLU-546	[-1, 1600, 16, 16]	0
Conv2d-547	[-1, 128, 16, 16]	204800
BatchNorm2d-548	[-1, 128, 16, 16]	256
ReLU-549	[-1, 128, 16, 16]	0
Conv2d-550	[-1, 32, 16, 16]	36864
BatchNorm2d-551	[-1, 1632, 16, 16]	3264
ReLU-552	[-1, 1632, 16, 16]	0
Conv2d-553	[-1, 128, 16, 16]	208896
BatchNorm2d-554	[-1, 128, 16, 16]	256
ReLU-555	[-1, 128, 16, 16]	0
Conv2d-556	[-1, 32, 16, 16]	36864
BatchNorm2d-557	[-1, 1664, 16, 16]	3328
ReLU-558	[-1, 1664, 16, 16]	0
Conv2d-559	[-1, 128, 16, 16]	212992
BatchNorm2d-560	[-1, 128, 16, 16]	256
ReLU-561	[-1, 128, 16, 16]	0
Conv2d-562	[-1, 32, 16, 16]	36864
BatchNorm2d-563	[-1, 1696, 16, 16]	3392
ReLU-564	[-1, 1696, 16, 16]	0
Conv2d-565	[-1, 128, 16, 16]	217088
BatchNorm2d-566	[-1, 128, 16, 16]	256
ReLU-567	[-1, 128, 16, 16]	0
Conv2d-568	[-1, 32, 16, 16]	36864
BatchNorm2d-569	[-1, 1728, 16, 16]	3456
ReLU-570	[-1, 1728, 16, 16]	0
Conv2d-571	[-1, 128, 16, 16]	221184
BatchNorm2d-572	[-1, 128, 16, 16]	256
ReLU-573	[-1, 128, 16, 16]	0
Conv2d-574	[-1, 32, 16, 16]	36864
BatchNorm2d-575	[-1, 1760, 16, 16]	3520
ReLU-576	[-1, 1760, 16, 16]	0
Conv2d-577	[-1, 128, 16, 16]	225280
BatchNorm2d-578	[-1, 128, 16, 16]	256
ReLU-579	[-1, 128, 16, 16]	0
Conv2d-580	[-1, 32, 16, 16]	36864
BatchNorm2d-581	[-1, 1792, 16, 16]	3584
ReLU-582	[-1, 1792, 16, 16]	0
Conv2d-583	[-1, 128, 16, 16]	229376
BatchNorm2d-584	[-1, 128, 16, 16]	256
ReLU-585	[-1, 128, 16, 16]	0
Conv2d-586	[-1, 32, 16, 16]	36864
BatchNorm2d-587	[-1, 1824, 16, 16]	3648
ReLU-588	[-1, 1824, 16, 16]	0
Conv2d-589	[-1, 128, 16, 16]	233472
BatchNorm2d-590	[-1, 128, 16, 16]	256
ReLU-591	[-1, 128, 16, 16]	0
Conv2d-592	[-1, 32, 16, 16]	36864
BatchNorm2d-593	[-1, 1856, 16, 16]	3712
ReLU-594	[-1, 1856, 16, 16]	0
Conv2d-595	[-1, 128, 16, 16]	237568
BatchNorm2d-596	[-1, 128, 16, 16]	256
ReLU-597	[-1, 128, 16, 16]	0
Conv2d-598	[-1, 32, 16, 16]	36864
BatchNorm2d-599	[-1, 1888, 16, 16]	3776
ReLU-600	[-1, 1888, 16, 16]	0
Conv2d-601	[-1, 128, 16, 16]	241664
BatchNorm2d-602	[-1, 128, 16, 16]	256
ReLU-603	[-1, 128, 16, 16]	0
Conv2d-604	[-1, 32, 16, 16]	36864
BatchNorm2d-605	[-1, 1920, 16, 16]	3840
AdaptiveMaxPool2d-606	[-1, 1920, 1, 1]	0
AdaptiveAvgPool2d-607	[-1, 1920, 1, 1]	0
AdaptiveConcatPool2d-608	[-1, 3840, 1, 1]	0
Flatten-609	[-1, 3840]	0
BatchNorm1d-610	[-1, 3840]	7680
Dropout-611	[-1, 3840]	0
Linear-612	[-1, 512]	1966592
ReLU-613	[-1, 512]	0
BatchNorm1d-614	[-1, 512]	1024
Dropout-615	[-1, 512]	0
Linear-616	[-1, 5]	2565

Total params: 20,070,789

Trainable params: 2,206,917

Non-trainable params: 17,863,872

----------------------------------------------------------------

Input size (MB): 3.00

Forward/backward pass size (MB): 2294.66

Params size (MB): 76.56

Estimated Total Size (MB): 2374.23

----------------------------------------------------------------

Results

First Stage

Second Stage

Third Stage

Confusion Matrix

References

1. What is Diabetic Retinopathy?
2. Blur Detection

Authors

• Abhik Sarkar
• Fastai

License

• Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License