The first Artificial Intelligence for Digital Pathology counted with ~ 30 participants that attended the Keynote from Prof. Nasir Rajpoot and 6 presentations from contributed papers. 
The Keynote speaker presented a thorough overview of the domain, describing concepts such as the difference between digital and computational pathology. In addition to results from the Tissue Analitics Lab in the University of Warwick, Prof. Nasir Rajpoot showed the potential of Computational Pathology in elucidating biological mechanisms, specifically in cancer research.
 
The topics of the contributed talks highlighted some of the challenges of AI in Digital Pathology. One of them is the availability of labelled data. The presentations by Geetank Raipuria and  Niccolo' Marini addressed the problem by using teacher/student paradigms, either to learn from noisy annotations or to generalize even in absence of local annotations. Interestingly, both presentations were applied to prostate cancer, which together with the presentation by Anjani Dhragadhariya completed the three talks dedicated to prostate cancer. In this case, the approach from Dhragadhariya, was based on performing NLP of the pathology reports to predict Gleason grade. One of the ideas that arose during the discussion among the participants was whether this additional source of data could be beneficial for the two other approaches that dealt with prostate cancer cases, given that both of them faced limitations in terms of annotations. 
The absence of local annotations was again discussed from the perspective of Multiple Instance Learning. In his talk, Ido Ben-Shaul proposed a method based on Montecarlo dropout to perform certainty-based pooling to classify whole slides based on the patches thought to carry more evidence of the disease.
 
On the topic of Stain Normalization, the talk by Yash Sharma proposed an cyclic adversarial mechanism to obtain a stain normalization that would keep the structural similarity. The discussion with Prof. Rajpoot pointed out that although stain normalization is effective from AI practitioners and allows better generalization of the models, it is often undesired by pathologists who see a potential information loss. The concept of stain augmentation was proposed to allow AI models to generalize to multiple stains variations, without modifying the original data. 
 
Finally, a thorough analysis of the data leakage phenomenon and its impact on evaluation metrics by Nicole Bussola, prompted the discussion on the publication of  overly optimistic results that cannot generalize to unseen datasets. The discussion evolved from the topic of how to perform proper evaluation of the methods to how to obtain demographically relevant datasets that allow AI models to generalize. In this regard, Prof. Rajpoot described how the Pathlake network could allow scientist to access relevant data.




Participants

AIDP2021-Screen


Program

International Workshop on Artificial Intelligence for Digital Pathology (AIDP 2021)

January 10, 2021

12:00 - 16:00

Program

Time line

Talk

Authors

12:00 - 12:15

Opening

Organizers

12:15 - 13:15

Invited Talk

Nasir Rajpoot, University of Warwick, UK.

13:15 - 13:35

Noise-Robust Training of Segmentation model using Knowledge Distillation

 

Geetank Raipuria, Saikiran Bonthu, and Nitin Singhal

13:40 - 14:00

Semi-supervised learning with a teacher-student paradigm for histopathology classification: a resource to face data heterogeneity and lack of local annotations

Niccolo Marini, Sebastian Otalora, Henning Muller, and Manfredo Atzori

14:05 - 14:25

Self-Attentive Adversarial Stain Normalization

Aman Shrivastava, William Adorno, Yash Sharma, Lubaina Ehsan, S. Asad Ali, Sean R. Moore, Beatrice Amadi, Paul Kelly, Sana Syed, and Donald E. Brown

14:30 - 14:50

Certainty Pooling for Multiple Instance Learning

 

Jacob Gildenblat, Ido Ben-Shaul, Zvi Lapp, and Eldad Klaiman

 

14:55 - 15:15

Classification of noisy free-text prostate cancer pathology reports using natural language processing

 

Anjani Dhrangadhariya, Sebastian Otalora, Manfredo Atzori, and Henning Muller

 

15:20 - 15:40

AI slipping on tiles: data leakage in digital pathology

 

Nicole Bussola, Alessia Marcolini, Valerio Maggio, Giuseppe Jurman, and Cesare Furlanello

 

15:40 - 16:00

Discussion and closing

Organizers



Materials

Presentations

Authors

Title

Link

 

Nasir Rajpoot, University of Warwick, UK.

Invited Talk

 

Geetank Raipuria, Saikiran Bonthu, and Nitin Singhal

Noise-Robust Training of Segmentation model using Knowledge Distillation

 

Paper 02

Niccolò Marini, Sebastian Otàlora, Henning Müller, and Manfredo Atzori

Semi-supervised learning with a teacher-student paradigm for histopathology classification: a resource to face data heterogeneity and lack of local annotations

Paper 03

Aman Shrivastava, William Adorno, Yash Sharma, Lubaina Ehsan, S. Asad Ali, Sean R. Moore, Beatrice Amadi, Paul Kelly, Sana Syed, and Donald E. Brown

Self-Attentive Adversarial Stain Normalization

Paper 04

Jacob Gildenblat, Ido Ben-Shaul, Zvi Lapp, and Eldad Klaiman

Certainty Pooling for Multiple Instance Learning

Paper 05

Anjani Dhrangadhariya, Sebastian Otálora, Manfredo Atzori, and Henning Müller

 

Classification of noisy free-text prostate cancer pathology reports using natural language processing

 

Paper 07

Nicole Bussola, Alessia Marcolini, Valerio Maggio, Giuseppe Jurman, and Cesare Furlanello

AI slipping on tiles: data leakage in digital pathology

Paper 11