What I Will Do In a different way Subsequent Time
Incorporate Grad-CAM for interpretability — must see the place these fashions focus when making selections. Preprocessing will also be improved; histopathological photographs typically profit from stain normalization and distinction enhancement. Whereas my dataset was comparatively balanced, delicate class imbalance may need affected DenseNet’s stability.
I can even fine-tune the present fashions and consider their efficiency on new, real-world samples. Superb-tuning with extra… particularly domestically acquired photographs may assist the fashions be taught the distinctive visible patterns and constraints present in low-resource medical settings . This can be a very essential step if these instruments are to be virtually deployed. I’ll undoubtedly share how that goes — proper right here.
Closing Ideas
This wasn’t nearly evaluating fashions. For me, it was extra about understanding how AI behaves within the messy, textured world of actual healthcare information.
ResNet50V2, a traditional, walked away as essentially the most dependable possibility. It proved that newer isn’t at all times higher and that’s particularly when the stakes are human lives.
And that, for me, was essentially the most precious shock of all.
Be taught Extra
For those who’re curious to dive deeper into our AI-powered telepathology pipeline, try this video the place we stroll by way of the challenge intimately — from microscope to mannequin:
🎥 Undertaking Overview on YouTube
You too can discover the total codebase, coaching logs, confusion matrices, and outcomes of the three fashions:
A Notice of Thanks
Earlier than I wrap up, I need to sincerely thank Dr. Daniel Memeu Maitethia, my challenge lead, for his unwavering steering and dedication all through this journey. (On a light-weight be aware:when you ever dare miss our weekly report conferences, Dr. Memeu doesn’t simply get disenchanted… he delivers that legendary ‘I anticipated higher’ look that would reboot your complete analysis motivation 😅).
I additionally owe a giant thank-you to Dr. Amos Chege, my machine studying lecturer. At first, I actually thought he was overloading us with too many hands-on assignments. However now, I’m extremely grateful. These assignments pushed me to analysis deeper, code smarter, and develop the boldness to experiment boldly.
To Ezekiel Otieno, thanks for at all times stepping in when the code acquired cussed and for capturing the precise histology photographs we later used to validate mannequin predictions. Your assist was an actual game-changer.
And at last, to all of the analysis assistants within the Meru College Physics lab, I appreaciate you. Your constructive power, the morale, the laughs, and the moments of brainstorming was breakthroughs. Additionally, your presence made even the lengthy hours really feel significant.
And naturally, heartfelt appreciation to Meru University of Science and Technology for fostering a supportive analysis surroundings the place concepts like these can develop and thrive.
