MTLAs, doctors and artificial intelligence at MLL – a good team!

We take artificial intelligence (AI) for granted these days and often use it in our everyday lives without noticing. But what does this mean for diagnostics – today and in the future? At MLL, artificial intelligence is used in various areas of diagnostics and is constantly being further developed. This not only saves time but will also facilitate the training of new staff in the future and expand opportunities for global collaboration. The aim is to provide patients with easier access to precise and rapid diagnoses and thus access to the best possible therapy. About one and a half years ago, we already reported here in the magazine about AI at MLL
  • so it's high time for an update on the latest developments and findings.

Cytomorphology

In the area of blood smear analysis, AI was developed in cooperation with AWS on the basis of a self-created training database that can generate differential blood counts within a few seconds. To this end, 300–500 cells per smear are photographed at high resolution and fully automatically with a MetaSystems scanner, stored as individual images, and classified by an algorithm. The median accuracy of this process is already 94% (Fig. 1). Lower values mainly occur with cell types that are difficult to differentiate even for experienced physicians and MTAs because they are in transitional stages of maturation. Critical-pathological cell types and those that cannot be determined with certainty can be marked by the AI and checked by doctors and MTAs. Since January 2021, blood smears of each case are scanned by AI and evaluated by humans and AI together. The comparison makes it possible for the clinical benefit to be assessed in a prospective study and enables AI to be included in routine practice (BELUGA study, NCT04466059). AI-based analysis of bone marrow cells is already part of current research and development as well.

Figure 1 from: Pohlkamp et al. Machine Learning (ML) Can Successfully Support Microscopic Differential Counts of Peripheral Blood Smears in a High Throughput Hematology Laboratory. Blood 2020; 136 (Supplement 1): 45–46.

Immunophenotyping

In the BELUGA study, both immunophenotyping data and cytomorphology data are collected. The AI for classifying lymphomas that has been developed by MLL in collaboration with AWS uses raw flow cytometric matrix data and allows for the prediction of diagnoses without prior manual visualization. Its use was able to be extended from mature B-cell neoplasms to other entities – the results are promising and of high precision, especially for multiple myeloma. One current challenge is the analysis of subpopulations within the sample, as is needed in myelodysplastic syndrome, for example. This could be remedied by an analysis where the existence of an entity is determined in a first step and subclassifications are studied by means of defined features in a second step.

Chromosome Analysis

The automatic generation of karyograms was already introduced into the routine at MLL in 2019. The AI for this has been steadily further developed, which has so far allowed the processing time of 1–3 minutes per karyogram (by very experienced staff) to be reduced to 31 seconds and a single mouse click. Copy number aberrations (CNAs) are automatically detected by the AI, and derivative chromosomes are prepared for manual assignment by a staff member. The precision for chromosome detection is 98.6%, and the majority of findings are made within 5 days (Fig. 2).

Figure 2 from: Haferlach et al. Artificial Intelligence Substantially Supports Chromosome Banding Analysis Maintaining Its Strengths in Hematologic Diagnostics Even in the Era of Newer Technologies. Blood 2020; 136 (Supplement 1): 47–48.

Molecular Genetics

Molecular genetics is playing an increasingly important role in diagnostics. Sequencing in particular gives deep insight into the biology of disease patterns. However, there are also new challenges. Not every mutation identified is relevant for the disease pattern. When assessing the pathogenicity, the MLL Predictor a specially developed AI, can assist the diagnostician significantly.

The next step in molecular genetic diagnostics is genome-wide approaches. However, the data generated by sequencing the complete genome and transcriptome exceeds the possibilities of manual analyses. This is where AI can help and increase the output of this data considerably. In collaboration with AWS MLL is developing an algorithm that uses 5 aberration types to make a diagnostic prediction. To this end, genetic structural variants, single nucleotide variants (SNVs), copy number variants (CNVs), gene fusions, and gene expression data are read into the tool and evaluated by the AI. The algorithm has high precision already, especially for genetically unique entities. Current training is increasingly focused on genetically less well-defined entities that pose a major challenge to the algorithm.

The author

»Do you have questions regarding this article or do you need further information? Please send me an e-mail.«

Dr. rer. nat. Constanze Kühn

Medical Writer