1 of 37

Classification of Gall Bladder Cancer Using Ensemble Learning and LSTM

BME 4000 Project/Thesis

Presented by,

Sourav Basak Shuvo

Roll: 1815010

4th Year, 2nd Semester

Dept. of Biomedical Engineering

Khulna University of Engineering & Technology

Supervisor

Dr. Mostafa Zaman Chowdhury

Professor

Dept. of Electrical and Electronic Engineering

Khulna University of Engineering & Technology, Khulna-9203

19 February, 2024

Page 1 of 37

Dept. of Biomedical Engineering, KUET

2 of 37

Content

Page number

Introduction

3-10

Literature Review

11

Proposed Methodology

12-19

Results

20-32

Discussion

33-34

Conclusion

35

References

36

Presentation Outline

19 February, 2024

Page 2 of 37

Dept. of Biomedical Engineering, KUET

3 of 37

Introduction

19 February, 2024

Page 3 of 37

Dept. of Biomedical Engineering, KUET

4 of 37

Background

  • Gallbladder Cancer
  • Neoplastic proliferation.
  • Prevalent biliary cancer.
  • 5th GI tract cancer.
  • Rare malignancy.
  • High incidence, mortality.
  • Rare early detection.

Fig. 1. Gallbladder.

Fig. 2. Normal gallbladder.

Fig. 3. Benign gallbladder.

Fig. 4. Malignant gallbladder.

Liver

Stomach

Bile

Gallbladder

19 February, 2024

Page 4 of 37

Dept. of Biomedical Engineering, KUET

5 of 37

Motivation

  • High mortality.
  • Ultrasound reliance.
  • Late diagnosis.
  • Interpretation challenges.
  • Diagnostic challenges.
  • Malignant complexity.
  • Early detection.
  • Automated classification.
  • Improve accuracy.

19 February, 2024

Page 5 of 37

Dept. of Biomedical Engineering, KUET

6 of 37

Problem Statement

  • Costly detection methods.
  • Noisy ultrasound images.
  • Sparse DNN research.
  • Low model accuracy.

19 February, 2024

Page 6 of 37

Dept. of Biomedical Engineering, KUET

7 of 37

Thesis Objectives

  • Review literature.
  • Analyze articles.
  • Study models.
  • Develop detection approach.
  • Introduce preprocessing pipeline.
  • Integrate LSTM.
  • Assess model performance.
  • Compare accuracies.

19 February, 2024

Page 7 of 37

Dept. of Biomedical Engineering, KUET

8 of 37

Contribution

  • Preprocessing pipeline.
  • LSTM integration.
  • Superior accuracy.
  • Early GBC detection.

19 February, 2024

Page 8 of 37

Dept. of Biomedical Engineering, KUET

9 of 37

Learning Outcomes

  • Acquired knowledge on
  • DL in medical imaging,
  • GBC classification & localization,
  • DNNs application,
  • Python, Keras, TensorFlow, & PyTorch,
  • Visio, Draw.io, Figma,
  • MathType usage, and
  • research writing & referencing,

19 February, 2024

Page 9 of 37

Dept. of Biomedical Engineering, KUET

10 of 37

Publication

S. B. Shuvo and M. Z. Chowdhury, “Classification of Gallbladder Cancer using Average Ensemble Learning,” in proc. of 6th International Conference on Electrical Engineering and Information & Communication Technology (ICEEICT), Dhaka, Bangladesh, 2024. [Under Review]

[1]

19 February, 2024

Page 10 of 37

Dept. of Biomedical Engineering, KUET

11 of 37

Literature Review

Ref.

Year

Main Contributions

[9]

2022

Development of GBCNet to identify GBC by utilizing a multi-scale, 2nd-order pooling architecture.

[10]

2022

The integration of hard negative mining is implemented in the creation of an unsupervised contrastive learning framework for the purpose of training image representations derived from ultrasound videos.

[11]

2023

Provides interpretable explanations consistent with medical literature.

[12]

2023

Demonstrated equivalent or superior diagnosis accuracy compared to skilled radiologists in identifying gallbladder cancer by ultrasound, even when stones, constricted gallbladders, tiny lesions, and neck lesions were present.

[13]

2023

Comparative analysis of the efficacy of deep learning models and a radiologist in distinguishing between XGC and GBC based on ultrasound pictures.

This Work

Development of a preprocessing pipeline for gallbladder ultrasound images and deployment of an ensemble model combined with LSTM network

Table 1. Summary of the recent relevant studies.

19 February, 2024

Page 11 of 37

Dept. of Biomedical Engineering, KUET

12 of 37

Proposed Methodology

19 February, 2024

Page 12 of 37

Dept. of Biomedical Engineering, KUET

13 of 37

Proposed Methodology (1/7)

Fig. 5. Gallbladder ultrasound images from GBCU dataset (a) normal, (b) benign and (c) malignant.

Dataset Name

Source

Classes

Number of samples

GBCU dataset

Basu et al.

[9]

3

1255 images:

432 normal,

558 benign, and

265 malignant

Table 2. Dataset statistics for GBCU dataset.

19 February, 2024

Page 13 of 37

Dept. of Biomedical Engineering, KUET

14 of 37

Proposed Methodology (2/7)

Fig. 6. Methodology of the proposed work.

19 February, 2024

Page 14 of 37

Dept. of Biomedical Engineering, KUET

15 of 37

Proposed Methodology (3/7)

Dataset Preprocessing

Fig. 7. Image preprocessing pipeline.

19 February, 2024

Page 15 of 37

Dept. of Biomedical Engineering, KUET

16 of 37

Proposed Methodology (4/7)

Fig. 8. Splitting GBCU dataset.

Fig. 9. Sample images from the dataset by class; 0 labeled (benign), 1 labeled (malignant), and 2 labeled (normal).

Dataset Splitting

Dataset Labeling

Dataset Augmentation

Fig. 10. Train dataset with augmentation.

19 February, 2024

Page 16 of 37

Dept. of Biomedical Engineering, KUET

17 of 37

Proposed Methodology (5/7)

Parameters

Values

Rotation range (in radians)

0.1

Random zoom

0.2

Hyperparameters

Value

Input Shape

128 x 128 x 3

Early stopping

true

Batch

32

Patience

10

Metrices

‘accuracy’

Loss Function

‘sparse_categorical_crossentropy’

Monitor

‘val_accuracy’

Optimizer

‘adamax’

Mode

max

Table 3. Parameters for data augmentation.

Table 4. Hyperparameter settings for proposed model.

19 February, 2024

Page 17 of 37

Dept. of Biomedical Engineering, KUET

18 of 37

Proposed Methodology (6/7)

Fig. 11. Model architecture of the proposed work.

19 February, 2024

Page 18 of 37

Dept. of Biomedical Engineering, KUET

19 of 37

Proposed Methodology (7/7)

Stages

Output Shape

VGG16 (vgg16_block5_conv3)

8 x 8 x 512

VGG19 (vgg19_block5_conv3)

8 x 8 x 512

ResNet50 (resnet50_conv5_block3_out)

4 x 4 x 2048

XceptionNet (xception_block14_sepconv2_act)

4 x 4 x 2048

LSTM Block 1

256

LSTM Block 2

256

Concatenate

512

Flatten

512

Dense

256

Dropout

256

Dense

128

Dropout

128

Dense

64

Dense

3

Table 5. The detailed structure of the proposed model.

19 February, 2024

Page 19 of 37

Dept. of Biomedical Engineering, KUET

20 of 37

Results

19 February, 2024

Page 20 of 37

Dept. of Biomedical Engineering, KUET

21 of 37

Results (1/12)

Fig. 12. Train and validation accuracy for 5-fold cross validation.

19 February, 2024

Page 21 of 37

Dept. of Biomedical Engineering, KUET

22 of 37

Results (2/12)

Fig. 13. Train and validation loss for 5-fold cross validation.

19 February, 2024

Page 22 of 37

Dept. of Biomedical Engineering, KUET

23 of 37

Results (3/12)

Fig. 14. Confusion matrix for all 5 folds (a) fold 1, (b) fold 2, (c) fold 3, (d) fold 4, and (e) fold 5.

(c)

(b)

(e)

(d)

(a)

19 February, 2024

Page 23 of 37

Dept. of Biomedical Engineering, KUET

24 of 37

Results (4/12)

Fig. 15. Train accuracy for different models.

19 February, 2024

Page 24 of 37

Dept. of Biomedical Engineering, KUET

25 of 37

Results (5/12)

Fig. 16. Validation accuracy for different models.

19 February, 2024

Page 25 of 37

Dept. of Biomedical Engineering, KUET

26 of 37

Results (6/12)

Fig. 17. Train loss for different models.

19 February, 2024

Page 26 of 37

Dept. of Biomedical Engineering, KUET

27 of 37

Results (7/12)

Fig. 18. Validation loss for different models.

19 February, 2024

Page 27 of 37

Dept. of Biomedical Engineering, KUET

28 of 37

Results (8/12)

Fig. 19. Comparison of ROC curve for class 0 (benign) between different models.

19 February, 2024

Page 28 of 37

Dept. of Biomedical Engineering, KUET

29 of 37

Results (9/12)

Fig. 20. Comparison of ROC curve for class 1 (malignant) between different models.

19 February, 2024

Page 29 of 37

Dept. of Biomedical Engineering, KUET

30 of 37

Results (10/12)

Fig. 21. Comparison of ROC curve for class 2 (normal) between different models.

19 February, 2024

Page 30 of 37

Dept. of Biomedical Engineering, KUET

31 of 37

Results (11/12)

Fig. 22. Bar plot for mean test accuracy & standard deviation between different models.

19 February, 2024

Page 31 of 37

Dept. of Biomedical Engineering, KUET

32 of 37

Results (12/12)

Fig. 23. Bar plot for performance matrices between different models.

19 February, 2024

Page 32 of 37

Dept. of Biomedical Engineering, KUET

33 of 37

Discussion (1/2)

Methods

Acc

Loss

F1-score

Sens.

PPV

Kappa

Spec.

Radiologist A [1]

70.0

-

-

70.7

-

-

87.3

Radiologist B [1]

68.3

-

-

73.2

-

-

81.1

GBCNet [1]

87.7

-

-

91.9

-

-

96.7

Radformer [3]

90.2

-

-

92.9

 -

 -

90.0

Our Work

99.37

1.78

99.52

99.64

99.40

99.22

99.69

Table 6. Comparison with other models with GBCU dataset (gallbladder cancer classification)

19 February, 2024

Page 33 of 37

Dept. of Biomedical Engineering, KUET

34 of 37

Discussion (2/2)

Fig. 24. Comparison with other models with GBCU dataset (gallbladder cancer classification).

19 February, 2024

Page 34 of 37

Dept. of Biomedical Engineering, KUET

35 of 37

Conclusion

  • Unique model proposal
  • High performance metrics
  • Superior accuracy, precision
  • Cost-effective implementation

19 February, 2024

Page 35 of 37

Dept. of Biomedical Engineering, KUET

36 of 37

References

S. Basu, M. Gupta, P. Rana, P. Gupta, and C. Arora, “Surpassing the Human Accuracy: Detecting Gallbladder Cancer from USG Images with Curriculum Learning,” in Proc. of CVF Conference on Computer Vision and Pattern Recognition (CVPR), New Orleans, LA, USA, Jun. 2022, pp. 20854–20864.

S. Basu, S. Singla, M. Gupta, P. Rana, P. Gupta, and C. Arora, “Unsupervised Contrastive Learning of Image Representations from Ultrasound Videos with Hard Negative Mining,” in Proc. of Medical Image Computing and Computer Assisted Intervention – MICCAI 2022, Cham, 2022, pp. 423–433.

S. Basu, M. Gupta, P. Rana, P. Gupta, and C. Arora, “RadFormer: Transformers with Global–local Attention for Interpretable and Accurate Gallbladder Cancer detection,” Medical Image Analysis, vol. 83, p. 102676, Jan. 2023.

P. Gupta et al., “Deep-learning Enabled Ultrasound Based Detection of Gallbladder Cancer in Northern India: A Prospective Diagnostic Study,” The Lancet Regional Health - Southeast Asia, Sep. 2023.

P. Gupta et al., “Deep-learning Models for Differentiation of Xanthogranulomatous Cholecystitis and Gallbladder Cancer on Ultrasound,” Indian J Gastroenterol, Dec. 2023.

[9]

[10]

[11]

[12]

[13]

19 February, 2024

Page 36 of 37

Dept. of Biomedical Engineering, KUET

37 of 37

Thank You

19 February, 2024

Page 37 of 37

Dept. of Biomedical Engineering, KUET