The following survey papers are (partially) quoted to illustrate the concepts of these papers:
Lin et al. Dual-path network with synergistic grouping loss and evidence driven risk stratification for whole slide cervical image analysis. MIA 2021.
Cheng, et al. Robust whole slide image analysis for cervical cancer screening using deep learning. Nature Communications 2021.
Zhu, et al. Hybrid AI-assistive diagnostic model permits rapid TBS classification of cervical liquid-based thin-layer cell smears. Nature Communications 2021.
Zhao, et al. Clinical-Inspired Cytological Whole Slide Image Screening with Just Slide-Level Labels. Axiv 2023.
Stegmüller, et al. Self-Supervised Learning-Based Cervical Cytology Diagnostics in Low-Data Regime and Low-Resource Setting. arXiv 2023.
arXiv, 2306.03407
Many thanks to the original authors.
My current research field is whole-slide images (WSI) processing, and this page will provide my thoughts for processing problems. This page would be served as my discovery logs and should not be treated seriously.
Sect. 0 Survey
Sect. 0.1 Cytology WSI Analysis Survey
In fact, due to the sparsity of its publicly available data, there are few discoveries in DL in cytology whole-slide images (WSI). Most of them can be categorized into the following subfields:
(A one-stage design and a two-stage design paper are redacted since the confidential request)
As we may see, current methods are limiting themselves in two-stage networks, i.e., first feed the patches into a detector, and then aggregate the attributes, and recently, several one-stage methods borrowing concepts from histopathology are on the trend. This blog is intended to dig fine-grained methods to explore MIL in cytology methods.
Sect. 0.2 Histopathology WSI Analysis Survey
(Neglected specific methods introduction. Namely, MAXMIL/MeanMIL/ABMIL/CLAM/DSMIL/TransMIL/HIPT/DTFD/PMIL/S4MIL/ICMIL/HEAT/PatchGCN/BCL/ILRA-MIL… etc.)
Comments:
1. Current field of histopathology WSI analysis contains lots of attention-based aggregators MIL methods.
There exists high class unbalance problem even within a single WSI.
3. (From our experiment) Current MIL methods exists biased attention distribution problem, meaning the core problem of MIL is building the accurate mapping of the malicious patches and the benign patches. Also, the highly class imbalance within a single WSI contributes to this degredation.
To the best of my knowledge, all those mentioned methods can be roughly divided into the following fields:
No method is discovering the coarse-to-fine relationship. i.e., finer level class has dependencies on coarser level class.
No MIL classification methods for fine-grained cytology classification.
For current project, I attempt to and have discovered Directions 3&4.
Sect 1.1 Failed Attempts
Current coarse-to-fine design seems to related to the term hierarchical classification. Three possible sub-directions are presented as follows:
Tree-based classifier/loss:
Ref. Kim, Sangwon, Jaeyeal Nam, and Byoung Chul Ko. "Vit-net: Interpretable vision transformers with neural tree decoder." International Conference on Machine Learning. PMLR, 2022.
b. Transformer Prompting:
Ref. Wang, Wenhao, et al. "TransHP: Image Classification with Hierarchical Prompting." Advances in Neural Information Processing Systems 36 (2023).
Neither does these methods work on the cytology dataset based on ABMIL backbone.
c. Utilize Classifiers in different hierarchies (Sort like ensemble?)
This direction works in my attempts.
Sect 2. Method
A pretty simple method is formed like this. By using coarse-grained labels as guidance for constrain, this proposed network can serve as a fine-grained MIL method for cervical WSI classification.
This method comprises of two stages:
Instance-level Hierarchy Learning
In this stage, the attention learned from each individual patches are forced to align through class hierarchy.
Bag-level Hierarchy Learning
In this stage, the class probabilities are aligned through class hierarchy.
To further improve the effectiveness of fine-grained classification, we employed supervised contrastive learning and dynamic weighting to balance two branches.
