A Universal Deep-Learning Model for Immunohistochemical Staining Across Diverse Cancer Types
Abstract
Tissue biopsy
analysis using immunohistochemistry (IHC) is crucial for diagnosing cancer and guiding treatment decisions. However, interpreting these visual data requires specialized pathologists and can be subjective and time-consuming. This study presents a Universal IHC (UIHC) deep learning model for automated cell classification and tumor proportion score (TPS) assessment, trained on a large and diverse dataset encompassing various cancer types and IHC staining patterns.
Introduction
Immunohistochemistry (IHC) stains provide essential information for diagnosing and treating cancer. Involves analyzing stained tissue samples, labeled for specific molecules. manual analysis of stained tissue for identifying the presence and extent of protein expression within the tumor. However,
These complex images require the expertise of trained pathologists, making the process time-consuming and prone to subjective variations.
To automate this process and enhance diagnostic accuracy, researchers have developed deep learning models for IHC
analysis. While these models often perform well on curated datasets, they usually lack generalizability to previously unseen staining types or cancer types.
In this study, we aimed to develop a Universal IHC
(UIHC) model, capable of accurately analyzing diverse IHC stains and various cancers. We trained our model on a vast dataset encompassing a wider range of cancers than traditional datasets. We evaluated performance
on multiple types. We evaluated performance across
multiple cohorts, including novel stains and cancer
types not encountered during training. Our results.
We found that the UIHC model
demonstrated superior performance across a variety of staining types and achieved remarkable accuracy in predicting TPS scores, highlighting its potential as a transformative tool for democratizing access to IHC analysis and accelerating cancer diagnostics.
Results.
Developing a Multi-cohort Training Set.
Our study involved developing a
multi-cohort training dataset, encompassing
various staining types.
Dataset Collection<- A total of 3046 whole slide images (WSIs) were collected, spanning several cancer types. including lung, urothelial, and breast cancer. $The dataset was divided into three subsets: training, tuning (validation), and testing
(Supplementary Table 1).
and lung cancer was used for internal validation and the
remaining dataset was used for external
validation. These sets encompassed diverse IHC, with target proteins including PD-L1 22C3 (regardless of cancer,
type data set including various immunostains for disease pan-cancer comparison and various
immunstain (Supplementary
We also included two additional.>
The 22C3 lung data set shared by contributed to the training of a model, alreadyUploaded images a development rumour.
Tumor with various immunostains for disease comparison.
Choosing the optimal model
Our citizen. *
**, data split:
We trained eight models: single-cohort
individuals.** To validate the effect of pooling
stain types and cancer
to improve performance. Then we tested the
performance<< of each $. Our results showed that the model trained on all cohorts (PH-LUB) prepared, understood by curated data sets generated. In this study, we
used eight models: single-cohort, **performance;
.
It achieves performance comparable to centroid.
Fine-Tuned
because
pulmonary, Results are shown in
**Concentration.
Patient**
A total of 3046 images. The dataset was divided into three
subsets, Togo training. We used these models to
complete the dataset**.
Today these
subsets were generated, which represent micro
, the
challenge is
What are the potential limitations or challenges in implementing UIHC in real-world clinical settings?
## A Deep Learning Breakthrough: AI Diagnoses Cancer From Tissue Samples
**Today we have Dr. Alex Reed, a leading researcher in computational pathology, here to discuss a groundbreaking new study that could revolutionize cancer diagnosis.**
**Welcome to the show, Dr. Alex Reed.**
**Dr. Alex Reed:** Thank you for having me.
**So, tell us about this new study. What did your team achieve?**
**Dr. Alex Reed:** We developed a universal deep learning model called UIHC, capable of accurately analyzing a wide range of immunohistochemistry (IHC) stains used in cancer diagnosis.
Essentially, UIHC uses artificial intelligence to “read” these complex tissue samples, identifying cancer cells and quantifying the extent of protein expression within tumors.
**That sounds exciting. Why is this such a big deal?**
**Dr. Alex Reed:** IHC analysis is currently a time-consuming and laborious process, requiring trained pathologists to interpret these intricate stained slides. This can be subjective and lead to variations in diagnoses.
Our UIHC model automates this process, making it faster, more objective, and potentially accessible to a wider range of healthcare providers.
**How is this model different from other AI tools used in cancer diagnostics?**
**Dr. Alex Reed:** While other models exist, they often struggle to generalize to new staining types or cancer types. We trained UIHC on a massive and diverse dataset encompassing various cancers and staining patterns, empowering it to handle a wider range of cases.
**Alex Reed:** Our results demonstrated UIHC’s superior performance across diverse IHC stains and its remarkable accuracy in predicting tumor proportion score (TPS), a crucial factor in determining treatment options.
**This is potentially groundbreaking news for cancer patients. What are the next steps for this technology?**
**Dr. Alex Reed:** We’re enthusiastic about translating our findings into clinical practice. Right now, we’re working on validating UIHC on even larger and more diverse patient cohorts.
We envision a future where this technology is readily available in hospitals and clinics worldwide, accelerating cancer diagnoses and ultimately improving patient outcomes.
