1 of 9

Pneumonia Detection System

Deepmindset

Xingyi Gao

Yiqian Guo

Yongning Li

UTS CRICOS 00099F

2 of 9

Motivation: Why Pneumonia Detection Matters

  • Pneumonia is one of the leading causes of death globally — especially in children and the elderly.
  • Early and accurate diagnosis is critical but limited by time, expertise, and resource constraints.
  • There is an urgent need for automated, scalable, and interpretable diagnostic tools in medical imaging.

UTS CRICOS 00099F

UTS CRICOS 00099F

3 of 9

AI for Medical Impact

  • Our project leverages artificial intelligence to assist clinicians in identifying pneumonia from chest X-ray images.
  • Using a combination of deep learning, ClearML automation, and a user-friendly interface, we aim to reduce diagnostic delay, improve accuracy, and offer scalable deployment in real-world settings.

UTS CRICOS 00099F

4 of 9

Contributions Made:�AI for Pneumonia Detection

We present an AI-driven medical imaging project focused on detecting pneumonia and related lung conditions from chest X-ray images.

​

COVID

​

Lung Opacity

​

Normal

​

Viral Pneumonia

​

ClearML Integration:

  • Dataset artifact tracking
  • Automated model training and tuning

Streamlit Cloud Deployment:

  • Integrated GUI for prediction and visualization

Jira & Confluence Usage：

  • Team project management

Lessons learned:

Composition of the dataset

5 of 9

Dataset Imbalanced & Noisy Data

Improvement of model accuracy

High Computational Cost of HPO

5

Delete this blue circle and click icon to insert your image. See instructions on Slide 18

Delete this blue circle and click icon to insert your image. See instructions on Slide 18

Key Challenges We Overcame

6 of 9

Model Evolution

6

Sprint 1

Sprint 2

Sprint 3

Structure: 2×Conv + MaxPooling → Flatten → Dense layers

Optimizer: Adam (lr = 0.0005)

DNN

Tuned Parameters:

batch_size: [32, 64]

learning_rate_stage2: [5e-6, 1e-5]

HPO-Optimized VGG16

Base model: VGG16

Stage 1: Freeze convoluti-onal layers, train classifier

Stage 2: Unfreeze top layers, fine-tune entire model

Fine-tuned VGG16

Model V1

82%

Accuracy

Model V2

90%

Accuracy

Model V3

91%

Accuracy

7 of 9

Pipline, HPO Tuned, Evaluation and GUI

Parameters Tuned Result

Workflow Pipline

GUI Interface on Streamlit Cloud

The Final model Evaluation

8 of 9

Product demonstration

https://deepmindset.streamlit.app/

9 of 9

Thank you