=Paper=
{{Paper
|id=Vol-2595/endoCV2020_Balasubramanian_et_al
|storemode=property
|title=Semantic Segmentation, Detection AND Localisation of Mucosal
Lesions from Gastrointestinal Endoscopic Images Using SUMNET
|pdfUrl=https://ceur-ws.org/Vol-2595/endoCV2020_paper_id_s18.pdf
|volume=Vol-2595
|authors=Velmurugan Balasubramanian,Rajiv Kumar,Sarasa Jyothsna Kamireddi,Rachana Sathish,Debdoot Sheet
|dblpUrl=https://dblp.org/rec/conf/isbi/Balasubramanian20
}}
==Semantic Segmentation, Detection AND Localisation of Mucosal
Lesions from Gastrointestinal Endoscopic Images Using SUMNET==
https://ceur-ws.org/Vol-2595/endoCV2020_paper_id_s18.pdf
SEMANTIC SEGMENTATION, DETECTION AND LOCALIZATION OF MUCOSAL
LESIONS FROM GASTROINTESTINAL ENDOSCOPIC IMAGES USING SUMNET
Velmurugan Balasubramanian1 Rajiv Kumar2 Sarasa Jyothsna Kamireddi3
Rachana Sathish4 Debdoot Sheet5
1
Velmurugan Balasubramanian, School of medical science and Technology, IIT, Kharagpur, India
2
Rajiv Kumar, Department of Chemical Engineering, IIT, Kharagpur, India
3
Sarasa Jyothsna Kamireddi, Department of Electrical Engineering, IIT, Kharagpur, India
4
Rachana Sathish, Department of Electrical Engineering, IIT, Kharagpur, India
5
Debdoot Sheet, Centre of excellence in AI, Department of Electrical Engineering, IIT, Kharagpur, India
1. METHOD
We trained a fully convolutional neural network based on
SUMNet [1] architecture, described in Fig.1, using Pytorch,
for segmentation, detection and localization of lesions in Gas-
trointestinal endoscopic images using 386 images from EDD
2020 dataset [2]. An 80:20 training-validation split was fol-
lowed with additional weights given to the under-represented Fig. 1: SUMNet architecture.
classes depending upon their overall frequency of occurrence.
We augmented the dataset with rotation, affine, scaling, pro- Class Name No of instances
jective and multi-crop transformations to accommodate for Barrett’s Esophagus 21
the variations caused due to scope positioning and augmented Suspicious 6
with variable brightness and HSV values to accommodate High grade dysplasia 9
for images enhanced with narrow-band imaging and variable Cancer 1
lighting conditions [3] [4]. We used the ADAM learning rate Polyp 30
optimizer and binary cross-entropy loss function for training.
SUMNet features (i) an encoder-decoder architecture with Table 1: Distribution of the detected classes in the test dataset
the pooling indices of encoder being passed to the corre-
sponding decoder upsampling layers, (ii) encoder having a
Fig. 2 shows examples of the semantic masks and the bound-
VGG11 like architecture pre-initialized with ImageNet pre-
ing boxes we obtained for each of the five classes. We were
trained weights and (iii) concatenation of activations of the
able to obtain a semantic segmentation score of 0.538 with a
encoder with that of the decoder, combining the features of
standard deviation of 0.35 in our test submission and a mean
segmentation networks for natural and biomedical images [1].
detection score of 0.16 with a standard deviation of 0.074.
2. RESULT
3. REFERENCES
Our model was able to obtain dice coefficients of 0.977,
[1] Sumanth Nandamuri, Debarghya China, Pabitra Mitra,
0.974, 0.986, 0.987, 0.961 and 0.545, 0.219, 0.172, 0.339,
and Debdoot Sheet. Sumnet: Fully convolutional model
0.573 on the training and validation sets for Barretts oe-
for fast segmentation of anatomical structures in ultra-
sophagus, suspicious, high-grade dysplasia, cancer and polyp
sound volumes. 2019 IEEE 16th International Sympo-
classes respectively. A class-wise distribution of the abnor-
sium on Biomedical Imaging (ISBI 2019), Apr 2019.
malities detected in the test dataset is shown in Table 1 and
Copyright c 2020 for this paper by its authors. Use permitted under [2] Sharib Ali, Noha Ghatwary, Barbara Braden, Dominique
Creative Commons License Attribution 4.0 International (CC BY 4.0). Lamarque, Adam Bailey, Stefano Realdon, Renato Can-
Fig. 2: Illustration of the original images (top) and their cor-
responding semantic masks (bottom).
nizzaro, Jens Rittscher, Christian Daul, and James East.
Endoscopy disease detection challenge 2020. arXiv
preprint arXiv:2003.03376, February 2020.
[3] Georg Wimmer, Andreas Uhl, and Andreas Vecsei. Eval-
uation of domain specific data augmentation techniques
for the classification of celiac disease using endoscopic
imagery. In 2017 IEEE 19th International Workshop
on Multimedia Signal Processing (MMSP), pages 1–6.
IEEE, 2017.
[4] Andrea Asperti and Claudio Mastronardo. The effective-
ness of data augmentation for detection of gastrointesti-
nal diseases from endoscopical images. arXiv preprint
arXiv:1712.03689, 2017.