Next generation
tumor boards
Predicting tumor chemosensitivity and relapse risk using Deep Learning.
Improved Tumor Diagnosis
Current challenges
Histology analysis is at the cornerstone of cancer treatment, enabling precise identification and characterization of tumors through meticulous microscopic examination. This essential process lays the foundation for effective treatment plans and diligent patient follow-up. However, it is not without challenges. Current morphological classifications, such as SBR grading and AJCC staging for breast cancer (see Valderrama et al., 2024), do not always correlate perfectly with treatment response, potentially affecting patient outcomes
Classical approaches
Innovative solutions are revolutionizing predictive diagnostics, such as next-generation DNA sequencing and the use of gene panels on tumors or liquid biopsies. These cutting-edge methods offer significant potential to improve treatment outcomes by providing more precise and personalized data (Smith et al., 2022; Johnson et al., 2023). However, they come with notable challenges: they are expensive, often costing several thousand euros per test (Norton et al., 2023), can be impractical due to limited available tumor material, and are still not without limitations in accuracy and reliability..
Our innovative solutions
We propose an innovative solution to enrich the diagnostic arsenal of physicians: using tissue analysis and tumor morphology to directly predict, from the image, the chemosensitivity of tumors and the risk of relapse. Our methodologies, based on Deep Learning, offer a number of advantages:
Accessibility
Can be performed on the basis of the initial diagnostic examination.
Economical
Just a few hundred euros.
Performance
Complete medical reporting to offer cutting-edge customized medicine from the day of diagnosis.
By integrating the latest technological advances and combining morphological analysis of tumors with Deep Learning algorithms, we offer a high-performance, affordable and always feasible solution for more accurate, personalized diagnosis. Join us in transforming diagnostic medicine and offering every patient the best possible treatment from day one.
* Breast Cancer Research, “The limitations of histopathology for predicting response to treatment and risk of recurrence in breast cancer,” 2022.
** Journal of Clinical Oncology, “Challenges in correlating histopathological features with treatment outcomes in breast cancer,” 2021.
Our technology
State-of-the-art Deep Learning
Ummon chara-prAIdict combines state-of-the-art proprietary algorithms based on weakly supervised and self supervised deep learning, combined with traditional machine learning and contextual molecular pathways, ensuring performances and fine-grained predictions for high precision diagnostics (Valderrama et al. 2024)
Robustness
Our algorithms undergo rigorous testing on extensive external cohorts to ensure their reliability and accuracy. They are evaluated for biological consistency, which involves predicting patient outcomes using different samples to account for tumor heterogeneity.
Additionally, our algorithms are assessed for technical variability by being tested on different scanners, confirming their scanner-agnostic capabilities and ensuring robust technical correlation. This comprehensive validation process underscores our commitment to delivering precise and consistent diagnostic solutions.
Expert control
Our solution's performance
Our solution currently targets breast cancer patients undergoing neoadjuvant chemotherapy. Our solution analyzes diagnostic biopsies to predict chemosensitivity, prognosis and risk of relapse. We tested our solution on an independent cohort of 640 patients that our algorithm had never seen, from a center independent of our learning cohorts.
The clinical model using the reference classification was not related to therapeutic response.
The clinical model using the reference classification was not related to therapeutic response.
| Methods | Clinical model (CM) | ||
|---|---|---|---|
| OR (95% CI) | AUC | P | |
| HER2+ | 3.63 (0.07-185.45) | 0.51 | 1 |
| HER2-/RH+ | 0.436 (0.0237-8.01) | 0.56 | 0.595 |
| TNBC | 2.03 (0.57-7.15) | 0.47 | 0.308 |
Clinical Model utilizes the molecular subtype (4 classes: HER2+/RH+, HER2-/RH+, HER2+/RH-, and TNBC), AJCC staging (11 classes: 0, I, IA, IB, IIA, IIB, IIIA, IIIB, IIIC, X, K), and SBR grade (3 classes: I, II, III) information.
Results on an external cohort of 640 patients.
| Methods | Whole Slide Image Model (WSIM) | ||
|---|---|---|---|
| OR (95% CI) | AUC | P | |
| HER2+ | 2.70 (1.08 - 6.76) | 0.67 | 0.0358 |
| HER2-/RH+ | 20.56 (1.14-371.74) | 0.87 | 0.00413 |
| TNBC | 3.02 (1.18-7.74) | 0.71 | 0.0206 |
Results on an external cohort of 640 patients.
We were able to test our solution on a cohort
- for HER2+ tumors: 3.63 more chances of detecting chemosensitive tumors..
- for TNBC: 2.03 more chances of detecting chemosensitive tumors.
- for luminal tumours (ER+/HER2-): 20.56 more chances of detecting the chemosensitive tumour.
compared with the standard classification.
These excellent results are the subject of further studies to further improve our performance and our level of clinical proof. We have also been able to initiate development programs in other pathologies, such as ovarian cancer and Hodgkin’s lymphoma.