Page 104 - AI for Good-Innovate for Impact

AI for Good-Innovate for Impact

(MTB) from patient samples [2], a process that can take weeks [3]. This delay hinders timely
treatment initiation and increases the risk of transmission.

The current gold standard for TB diagnosis, culturing MTB especially in developing countries,
suffers from several limitations. It is time-consuming, labour-intensive, and requires specialized
facilities. Additionally, some strains of MTB grow slowly [4], further extending the diagnostic
window. This delay translates to prolonged patient suffering, increased healthcare costs,
and the potential for further transmission within the community. More importantly, in most
developing countries, no modern method is used, therefore leading to severe outcomes like
damaged lungs.

There is a faster alternative called sputum smear microscopy that uses Ziehl-Neelsen (ZN)
staining to identify acid-fast bacilli (AFB) characteristic of MTB. However, it has low sensitivity
[5], particularly in patients with low bacterial loads; therefore, it is not always reliable, especially
if the sickness is in its early stages. Additionally, differentiating MTB from other acid-fast bacilli
can be challenging for less experienced technicians. This signifies a need for a better approach
to diagnosis.

We propose the use of multiple deep learning techniques (image segmentation and classification
with state-of-the-art architectures) to analyze microscopic images of patient samples stained
for MTB. AI has mostly been used to diagnose TB through the classification of conventional
radiography images [6]. This approach has shown a good change, but since it depends on
visible lung damage, there is usually a delay in realising the infection.

A combination of trained model architectures could potentially identify MTB bacilli with high
accuracy and speed, significantly reducing diagnostic turnaround time, therefore bringing the
following benefits.

• Faster treatment initiation will allow for early diagnosis and prompt treatment, minimizing
patient complications, worsening, and transmission risks.
• Building this will lead to improved resource allocation, if we don’t depend on culturing,
this frees up resources for other critical laboratory procedures.
• AI powered diagnosis could be deployed in resource-limited settings lacking specialized
facilities for culturing. This is more important in developing countries, where most health
centers lack specialized facilities.
The possibility of drawbacks is more centered on the quality and size of the data to be used in
building the model, this mostly affects the model’s accuracy. Additionally, integrating AI into
clinical workflows requires careful validation and regulatory approval.

UN Goals:

SDG 3: Good Health and Well-being

Justify UN Goals selection: Tuberculosis remains a significant barrier to achieving SDG
3, particularly target 3.3. It states that "end the epidemics of AIDS, tuberculosis, malaria
and neglected tropical diseases and combat hepatitis, water-borne diseases and other
communicable diseases" and the targets linked to the end TB strategy are two, one is to detect
100% of new sputum smear-positive TB cases, cure at least 85% of these cases, and eliminate TB
as a public health problem (<1 case per million population) by 2050. Therefore, this proposed
solution directly faces SDG 3 and particularly 3.3, as explained. Faster diagnosis through AI
leads to quicker initiation of treatment, improving patient outcomes, and reducing the risk of

88