Tjalling Bosse

LUMC

In short, Tjalling Bosse's research entails the following: 

AIR-MEC: Artificial Intelligence to Refine the Molecular Endometrial Cancer Classification

'Endometrial (womb) cancer is the most common gynaecological malignancy in the Western world, affecting 100,000 women each year in Europe alone. While most women with endometrial cancers are cured by treatment, our inability to identify when surgery alone is sufficient means that many patients receive radiotherapy or chemotherapy that they did not need. For the appreciable fraction of women whose disease has spread at the time of diagnosis, or returns after surgery, the prognosis is poor and treatment options are limited.

Predicting spread or recurrence of endometrial cancer following surgery, and decisions to give additional treatments, have traditionally been based on the appearance of the cancer under a microscope. My previous work has shown that these predictions can be improved by also doing molecular (genetic) tumor testing. This helps to determine the specific type of endometrial cancer and predict how it will behave; however, such molecular testing is costly and not available everywhere.

Artificial intelligence could provide additional information to that which we currently get from molecular testing, and in hospitals where molecular testing is unavailable, serve as an alternative. In this project we will combine image analysis with molecular testing results from participants with womb cancer from the PORTEC trials. This represents an extremely powerful dataset and is one of the world’s largest collections of endometrial cancers. AIR-MEC should help us to better predict outcomes for patients depending on their specific endometrial cancer type and tailor treatments accordingly.'

Brief summary of progress / results November 2023

The AIRMEC project is moving forward as planned. The team has swiftly identified a highly skilled PhD candidate (Sarah Fremond) to lead AIRMECs experiments. She has now digitalized > 3000 unique endometrial cancer slides from the PORTEC tissue repository. Furthermore, AIRMEC team has been able to build a first version of a molecular prediction model that is capable of predicting molecular class based one H&E image. The current performance of this model is state-of-the art and novel insights into imagefeatures that characterize specific tumors have already been identified. This work has already reached international appreciation, and was awarded for a platform presentation at the USCAP conference, in Los Angeles, USA (March 2022). The work was published in the Lancet Digital Health in february 2023. Subsequently the team has focused on developing a model that is capable of accuratly predict risk of distant recurrence from one tumour containing H&E slide. This model, called HECTOR, has a innovative mulimodal design, which incorporates the previous molecular classfication prediction model. HECTOR outperforms the current gold standard, without the need for expensive molecular testing. HECTOR implementation has the ability to reduce over- and undertreatment, by better identifying which patients are at low- or high-risk of recurrence. This work has been recognized by the AACR 2023 annual conference, where it was selected for presentation at the press conference. The manuscript descring the final performance of HECTOR is currently being prepared for submission.

Brief summary of progress / results December 2022

The AIRMEC team has achieved major milestones in their project this year, by finalizing and publishing their first deep learning model that can predict molecular classes using one histology slide containg endometrial cancer (im4MEC). In 2022 the PhD candidate on the project (Sarah Fremond) first presented the initial results for an international pathology audience at the USCAP conference, in Los Angeles, USA (March 2022). Later in the year (October 2022) the work also got international attention at the ESGO conference in Berlin (Germany) for a more clinical audience. This work, with its state-of-the-art model performance and  interpretable results, suggests that a deep learning based model can have a relevant role in endometrial cancer diagnostics. The identification of morpho-molecular correlates and their impact on prognosis advances the evidence towards building an improved risk stratification system in endometrial cancer and unifying the moleculardriven and morphology-driven classification systems. Future work should focus on prospective validation, the further exploration of tumour heterogeneity in endometrial cancer, solving the current difficulty in distinguishing POLEmut from mismatch repair deficient endometrial cancer, and reproducibility on biopsy specimens. The AIRMEC team is proud to announce that flagship journal Lancet Digital Health has accepted the work on this model for publication. 

Brief summary of progress / results July 2022

The AIRMEC project is moving forward as planned. The team has swiftly identified a highly skilled PhD candidate (Sarah Fremond) to lead AIRMECs experiments. She has reached the first milestone by having digitalized >2000 unique endometrial cancer slides from the PORTEC tissue repository. Furthermore, AIRMEC team has been able to build a first version of a molecular prediction model that is capable of predicting molecular class based one H&E image. The current performance of this model is state-of-the art and novel insights into image-features that characterize specific tumors have already been identified. This work has already reached international appreciation, and was awarded for a platform presentation at the USCAP conference, in Los Angeles, USA (March 2022). The AIRMEC team will continue to improve the model’s performance and will expects to come out with an even better molecular prediction
model late 2022.

Publications 2022

The AIRMEC-team led by dr. Tjalling Bosse (pathologists at the Leiden University Medical Center) has developed an deep-learning model which can robustly predict molecular classes of endometrial carcinoma using routine H&E microscopy images. This is the first deep-learning showing state-of-the-art performance for such a complex task. An important element in this study is that the model (called “im4MEC”) is interpretable by design. This way, the research team was able to study the visual aspects that im4MEC uses to makes its predictions. The work is now published in the Lancet Digital Health and the AIRMEC-team will continue to build further on more advanced models aiming to improve the care of endometrial carcinoma patients. This work has been funded by the Hanarth Fonds. More information can be found here.