Software

Vision AI

At Microtechnix, we are committed to advancing the field of microbiology quality control (QC) through innovative, reliable technologies. Our cutting-edge AI solutions are designed to enhance the accuracy, efficiency, and compliance of your QC processes.

One of the key features of our AI-powered systems is the use of locked state AI. This approach ensures that once the AI’s performance has been rigorously validated, it remains in a secure, unchangeable state, guaranteeing consistent results and full compliance with regulatory standards

The Power of Locked State AI

Locked state AI means that once the AI model’s performance is verified and validated in real-world conditions, it is then securely “locked,” ensuring that its behavior and decision-making processes cannot be altered. This guarantees:

Consistency: After validation, the AI will perform the same way every time, providing repeatable and reliable results.
Regulatory Compliance: By locking the AI, we meet stringent 21 CFR Part 11 compliance requirements, ensuring all data processing is traceable, auditable, and unalterable once validated.
Transparency: All decisions made by the AI are fully traceable, with a clear audit trail. This means you can always track when and how the AI made its determinations.
Security: Locked state AI ensures that after the system has been validated, no unauthorized modifications can be made to the AI algorithms, preserving data integrity.

Why vision AI?

Traditional computer vision solutions comes with several challenges to efficiently automate image analysis, especially when shape and size of objects differ a lot.

Here’s why Vision AI can positively impact computer vision projects:

‍Faster image labelling
Full traceability & transparency
The ability to resolve much complexer use cases
Scalable & Robust

At Microtechnix, 2 vision AI template models are available for immediate deployment at customers:

EMdi assistant: detects colony growth on plates based on a binary classification model
EMdi automation: automatically counts colonies on positive plates based on an object detection model

These AI models can, on request, be re-trained with customer-specific data to further optimize performance characteristics to meet unique requirements. We provide a user-friendly environment for labeling images that can be used to re-train the existing models, enhancing their accuracy for specific tasks and environments. Once the model is deployed, it enters a locked-state. This means that the model’s parameters are fixed and cannot be altered without proper authorization, ensuring data integrity, security, and compliance with regulatory standards. This locked-state provides confidence that the AI’s behavior remains consistent and reliable throughout its usage, without unauthorized modifications.

Building a representative dataset

Essential in the overall development of a performant vision AI model is the collection of high quality data, subsequent analysis and finally the validation of the data.

The dataset that supports the pre-trained models include balanced data buckets from :

Negative plates (no colony)
Plates containing defects (writing, crack, injection point, etc.)
Plates with fogging (light condensation)
Positive plates (bacteria present)
Positive plates (mold present)

This results in a highly diverse dataset representing a real world environment.

Data Labeling

Data labelling was performed by in-house imaging and microbiology experts to ensure correct labels are applied across the entire dataset. By ensuring microbiology experts and imaging experts agree on the data label, the 4-eyes principle was applied to guarantee the highest quality labelled dataset.

AI training

The developed AI-models were trained on real-world representative datasets.

The key performance objectives are:

No false negatives
<5% false positives
Counting accuracy exceeding 98% for Petri dishes with less than 100 CFUs
Able to run on a local computer (edge device)

Validation

Datasets are split into 3 categories:

Train
Validate
Test

To ensure proper evaluation of the AI model performance. A balanced split of data buckets is made, enabling performance validation on large numbers.

How can end users benefit from AI solutions?

Image analysis is a time intensive process and is prone to human errors. Assisted or automated diagnostics enables more accurate analysis at a fraction of the time and cost of traditional methods.
‍

Implementing AI offers various benefits: