 ICACT20240372 Slide.00
[Big Slide]
|
Chrome  Click!! |
.jpg) |
Overview of the potentials of multiple instance learning in cancer diagnosis: Applications, challenges, and future directions.
Tagne Poupi Theodore Armand Institute of Digital Anti-Aging Healthcare Inje University Republic of Korea Supervisor: Professor Hee-Cheol Kim
|
| ICACT20240372 Slide.01
[Big Slide]
|
Chrome Click!! |
.jpg) |
Index's |
| ICACT20240372 Slide.02
[Big Slide]
|
Chrome Click!! |
.jpg) |
Introduction.
Despite advances in cancer research, the mortality rate of cancer patients is still alarming, rendering it a severe threat to human life
Accurate and timely cancer diagnosis is crucial in determining appropriate treatment strategies and improving patient survival rates
Traditional diagnostic methods, such as histopathological analysis, imaging techniques, and biomarker assessments, rely on subjective interpretations and may have limitations in terms of sensitivity, specificity, and reproducibility
In recent years, advanced machine-learning techniques have emerged as powerful tools to address challenges and enhance cancer diagnosis
|
| ICACT20240372 Slide.03
[Big Slide]
|
Chrome Click!! |
.jpg) |
Introduction
While using Computer-aided diagnosis (CAD) and machine learning methods, cancer datasets often exhibit characteristics such as class imbalance, noisy labels, and complex interactions between features, further complicating accurate diagnosis.
Overcoming these challenges enables a growing need for advanced machine-learning techniques in cancer diagnosis
By leveraging computational power and advanced algorithms, machine-learning approaches can improve the sensitivity and specificity of cancer diagnosis, enhance risk stratification, and aid in treatment decision-making.
|
| ICACT20240372 Slide.04
[Big Slide]
|
Chrome Click!! |
.jpg) |
Multiple instance learning (MIL) is a machine learning paradigm that has gained attention for its applicability in many areas, including image classification, drug activity prediction, text classification, and object recognition in computer vision, among others
MIL is a weakly supervised learning approach that deals with specific types of problems and is helpful in cases where the supervised methods are unsuitable because of the lack of colossal labels or annotations in the dataset.
In MIL, training data is organized into bags with instances.
|
| ICACT20240372 Slide.05
[Big Slide]
|
Chrome Click!! |
.jpg) |
Multiple instance learning (MIL) is a machine learning paradigm that has gained attention for its applicability in many areas, including image classification, drug activity prediction, text classification, and object recognition in computer vision, among others
MIL is a weakly supervised learning approach that deals with specific types of problems and is helpful in cases where the supervised methods are unsuitable because of the lack of colossal labels or annotations in the dataset.
In MIL, training data is organized into bags with instances.
An instance can be either positive (belonging to a specific target class) or negative (not belonging to the target class).
|
| ICACT20240372 Slide.06
[Big Slide]
|
Chrome Click!! |
.jpg) |
Figure |
| ICACT20240372 Slide.07
[Big Slide]
|
Chrome Click!! |
.jpg) |
Fundamental of multiple instance learning
Multiple Instance Learning (MIL) is gaining interest from industry and academia because it fits a variety of problems and allows us to leverage weakly labeled data
Multiple Instance Learning (MIL) is a form of weakly supervised learning where training instances are arranged in sets called bags, and a label is provided for the entire bag
In MIL, training data is organized into bags with instances.
An instance can be either positive (belonging to a specific target class) or negative (not belonging to the target class).
|
| ICACT20240372 Slide.08
[Big Slide]
|
Chrome Click!! |
.jpg) |
A whole slide image histopathological image can be represented as a bag of instances and its patches as instances using MIL
A Bag could represent a region of interest in an image, a patient's medical record, or a collection of molecular features
MIL uses bag-level labels, which can be either positive or negative.
MIL algorithms are usually divided into two main parts: An instance-level classifier and a bag-level classifier
|
| ICACT20240372 Slide.09
[Big Slide]
|
Chrome Click!! |
.jpg) |
An instance-level classifier provides information about the individual instances within each bag
The instance-level classifiers help to access the features of instances within each bag, and their training results in differentiating between positive and negative instances without directly using bag-level classification
Implementing instance-level classification uses various algorithms such as SVM, logistic regression, or k-nearest neighbors.
|
| ICACT20240372 Slide.10
[Big Slide]
|
Chrome Click!! |
.jpg) |
A bag-level classifier is responsible for making predictions at the bag level based on the collective information of the instances.
The output of the bag-level classifier is typically a probability or confidence score indicating the likelihood of a bag being positive or negative.
MIL uses majority voting, average probability, max probability, and clustering techniques to achieve bag-level classification
|
| ICACT20240372 Slide.11
[Big Slide]
|
Chrome Click!! |
.jpg) |
Figure |
| ICACT20240372 Slide.12
[Big Slide]
|
Chrome Click!! |
.jpg) |
A. Breast Cancer Diagnosis using MIL
Breast cancer diagnosis tasks, including subtype classification, mammogram analysis, and lesion detection.
B. Lung Cancer Diagnosis using MIL
MIL techniques have been applied to lung cancer diagnosis tasks, focusing on nodule detection and risk prediction.
C. Prostate Cancer Diagnosis using MIL.
MIL techniques have been employed in prostate cancer diagnosis, particularly in histopathology image analysis and detection in multiparametric MRI
|
| ICACT20240372 Slide.13
[Big Slide]
|
Chrome Click!! |
.jpg) |
Applications of multiple instance learning in cancer diagnosis |
| ICACT20240372 Slide.14
[Big Slide]
|
Chrome Click!! |
.jpg) |
Advantages of MIL in cancer diagnosis.
Reduced reliance on precise instance-level labels.
Accommodation of inter-instance variability
Robustness to noisy or mislabeled instances
|
| ICACT20240372 Slide.15
[Big Slide]
|
Chrome Click!! |
.jpg) |
Challenges associated with MIL.
Model interpretability
Bag-level feature representation
To overcome the limitations and improve the performance of MIL-based cancer diagnostic models, several strategies can be employed:
Instance selection or weighting: Selecting informative instances within bags or assigning instance weights can enhance the discrimination between positive and negative bags
Ensembling and boosting: Ensembling multiple MIL models or applying boosting techniques can enhance the overall performance and robustness of the models |
| ICACT20240372 Slide.16
[Big Slide]
|
Chrome Click!! |
.jpg) |
Prospects.
Integration of MIL with deep learning for enhanced cancer diagnostic accuracy
Exploration of MIL in other cancer types and clinical domains
Incorporating domain knowledge and expert annotations into MIL frameworks. |
| ICACT20240372 Slide.17
[Big Slide]
|
Chrome Click!! |
.jpg) |
Conclusion |
| ICACT20240372 Slide.18
[Big Slide]
|
Chrome Click!! |
.jpg) |
Thank You for Your Attention
|
