Scribd
Upload
21 views3 pages

Untitled Document

The document outlines future work in federated learning (FL) for medical image analysis, including real hospital tests, exploring various FL methods, and enhancing privacy while using more data. It also addresses limitations such as the lack of variety in images, testing on a single GPU, and the need for segmentation and rare-case detection. The conclusion highlights high accuracy achieved, a solid pipeline established, and a roadmap for further research in FL, personalization, and privacy.

Uploaded by

jobaerislam16
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
21 views3 pages

Untitled Document

The document outlines future work in federated learning (FL) for medical image analysis, including real hospital tests, exploring various FL methods, and enhancing privacy while using more data. It also addresses limitations such as the lack of variety in images, testing on a single GPU, and the need for segmentation and rare-case detection. The conclusion highlights high accuracy achieved, a solid pipeline established, and a roadmap for further research in FL, personalization, and privacy.

Uploaded by

jobaerislam16
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Slide: Future Work

Bullets

●​ Real Hospital Tests​

○​ Run training on separate hospital servers without sharing raw images​

●​ Try More FL Methods​

○​ Compare FedAvg, FedProx, FedBN and a personalized FL version​

●​ New Tasks​

○​ Teach the model to mark tumor areas (segmentation)​

○​ Spot rare or unusual cases automatically​

●​ Stronger Privacy​

○​ Add methods so updates don’t reveal patient details​

●​ Use More Data​

○​ Combine images with patient info (age, gender, stage) in FL​

Speaker Notes

“We will move from our single-computer tests to real hospital servers, each keeping
its own images private. We’ll try different federated methods—FedAvg, FedProx,
FedBN, and a personalized variant—to see which works best. Then we’ll expand
beyond simply classifying images by training the model to outline tumors and catch
rare cases. To protect privacy, we’ll add techniques so no one can reverse-engineer
patient data from updates. Finally, we’ll bring in basic patient info alongside the
images to train on both kinds of data together.”
Slide: Limitations
Bullets

●​ Limited Variety​

○​ 25 000 images all come from just 1 250 originals, so they look very similar​

●​ One-Box Testing​

○​ We ran everything on one GPU, not across real hospital networks​

●​ Single Model Used​

○​ We only tested MobileNetV2, not other architectures​

●​ Only Classification​

○​ We didn’t do segmentation or rare-case detection yet​

●​ Scaling Unknown​

○​ We don’t know how this setup will handle dozens of hospitals​

Speaker Notes

“We recognize that our images are all variations of 1 250 originals, so we may miss
some tissue differences. All our experiments ran on one GPU, not on real-world
hospital networks, so we didn’t face network delays or servers dropping out. We
only used MobileNetV2 and focused on classifying images—we haven’t yet done
tumor outlining or rare-case spotting. Finally, we don’t yet know how well this will
scale if dozens of hospitals join.”

Slide: Conclusion
Bullets

➔​ High Accuracy​

◆​ Our model hits about 99.7 % on five tissue classes​


➔​ Solid Pipeline​

◆​ We set up 70/15/15 data splits, EarlyStopping, and learning-rate control​

➔​ Research Gap Filled​

◆​ Nobody’s applied FL to LC25000 before—this is our baseline​

➔​ Clear Next Steps​

◆​ We have a roadmap for real FL, personalization, fairness, and privacy​

Speaker Notes

“In summary, we trained MobileNetV2 on the LC25000 dataset and achieved nearly
99.7 % accuracy across five tissue types. We built a full pipeline with a 70/15/15
split, early stopping, and learning-rate scheduling. A literature review showed few
prior federated-learning work on this dataset, so we’ve decided to go with this. We
also laid out clear next steps for real multi-site FL, model personalization, fair
performance across hospitals, and strong privacy protections.”

You might also like