1 of 14

[T-33]

Exploring the value of relabelling for multi-organ tumor segmentation from CT scans

Dzvinka Yarish

Nikita Fordui

Donatas Vaiciukevičius

Yuliia Siur

Project owner: Dmytro Fishman

2 of 14

Problem

Availability of annotated data is a

critical bottleneck when training

successful models.

Time of radiologists is expensive.

Hypothesis: data with artificially generated labels can still be useful

Images: https://cmescience.com/guide-to-becoming-a-radiologist/, https://www.healthline.com/health/kidney-health/ct-scan-for-kidney-stones, https://www.meme-arsenal.com/en/create/meme/917769

3 of 14

Datasets

KITS-19	DeepLesion
206 scans in total 164 - 42 Train/Test split Professional 3D annotations (segmentation masks) for kidneys and tumors for each scan.	356 scans with 495 lesions in total 171 - 185 Train/Test split 2D bounding boxes for tumors for one “key slice” per tumour. No kidney annotations (segmentation masks).

4 of 14

Model

https://github.com/MIC-DKFZ/nnUNet/tree/master/nnunet

Predictions are made with the KITS pre-trained nnUNet model.
nnUNet is a deep learning-based medical image segmentation framework

Image: https://medium.com/miccai-educational-initiative/nnu-net-the-no-new-unet-for-automatic-segmentation-8d655f3f6d2a

5 of 14

Our approach

nnUnet KITS+DL

nnUnet KITS only

KITS19

DeepLesion

1. train

2. predict test set

4. train

5. predict test set

3. get segmentations

6. predict test set

4. train*

* only use scans with Dice score between bounding box and segmentation > 0

use Dice score for evaluation

use manual inspection for evaluation

6 of 14

Results: Metrics

Training data	KITS-19 test set Dice score	DeepLesion kidney tumor detection proportion
KITS-19	0.833	171/356
KITS-19 + DeepLesion	0.852	69/185

Scans where no tumors were found by KITS only model

7 of 14

Results: KITS19 test set

KITS+DL model > KITS only model

8 of 14

Results: KITS19 test set

KITS+DL model < KITS only model

9 of 14

Results: DeepLesion test set

10 of 14

Results: DeepLesion test set

11 of 14

Results: DeepLesion test set

12 of 14

Conclusions

Adding artificially labeled data results in slight increase in model’s performance in kidney tumour detection
Working with 3D data is highly resource-intensive
nnUnet framework significantly simplifies the training process.

13 of 14

Thank you for attention

14 of 14

Our approach

Test pretrained KITS model on KITS-19 test set.
Get artificial segmentation masks for DeepLesion with pretrained on KITS-19 full resolution 3D nnUnet model.
Calculate Dice score between artificial masks and bounding boxes.
Add to train set DeepLesion scans with Dice score > 0.
Train 3D low resolution nnUnet model on mix of KITS-19 and DeepLesion data.
Train 3D low resolution nnUnet model on KITS-19 only for further comparison.
Test KITS-19 only and KITS-19+DeepLesion trained models on test set for KITS-19 and compare results.