This thesis explores the performance limits of semantic segmentation architectures, specifically applied to the detection of intestinal polyps in endoscopic images. Starting from advanced models (HSNet and PolypPVT) and ensemble techniques based on the Segment Anything Model (SAM), the study aims to quantify the maximum theoretical improvement margin (Upper Bound) through the introduction of "Oracles". These ideal algorithms (Oracle1, OracleW, OracleBox, OracleBoxes), granted access to the Ground Truth, guide the aggregation of predictive masks, highlighting how standard protocols have now saturated their learning capacity (with growth margins limited to a few percentage points). However, the results demonstrate that the true bottleneck lies in the training data: by violating the standard protocol and significantly expanding the training set (from 1,450 to 6,091 images) over a larger number of epochs, it is possible to unlock the untapped potential of the networks.
Il presente lavoro di tesi esplora i limiti prestazionali delle architetture di segmentazione semantica, specificamente applicate al riconoscimento di polipi intestinali in immagini endoscopiche. Partendo da modelli avanzati (HSNet e PolypPVT) e tecniche di ensemble basate sul Segment Anything Model (SAM), lo studio mira a quantificare il margine di miglioramento teorico massimo (Upper Bound) tramite l'introduzione di "Oracoli". Questi algoritmi ideali (Oracle1, OracleW, OracleBox, OracleBoxes), dotati di accesso alla Ground Truth, guidano l'aggregazione delle maschere predittive, evidenziando come i protocolli standard abbiano ormai saturato la loro capacità di apprendimento (con margini di crescita limitati a pochi punti percentuali). I risultati dimostrano tuttavia che il vero collo di bottiglia risiede nei dati di addestramento: violando il protocollo standard ed espandendo significativamente il training set (da 1.450 a 6.091 immagini) per un numero maggiore di epoche, è possibile sbloccare il potenziale inespresso delle reti.
Ensemble per segmentazione semantica: analisi sperimentale di limiti superiori alle prestazioni
SULOZEQI, SERGIO
2025/2026
Abstract
This thesis explores the performance limits of semantic segmentation architectures, specifically applied to the detection of intestinal polyps in endoscopic images. Starting from advanced models (HSNet and PolypPVT) and ensemble techniques based on the Segment Anything Model (SAM), the study aims to quantify the maximum theoretical improvement margin (Upper Bound) through the introduction of "Oracles". These ideal algorithms (Oracle1, OracleW, OracleBox, OracleBoxes), granted access to the Ground Truth, guide the aggregation of predictive masks, highlighting how standard protocols have now saturated their learning capacity (with growth margins limited to a few percentage points). However, the results demonstrate that the true bottleneck lies in the training data: by violating the standard protocol and significantly expanding the training set (from 1,450 to 6,091 images) over a larger number of epochs, it is possible to unlock the untapped potential of the networks.| File | Dimensione | Formato | |
|---|---|---|---|
|
Sulozeqi_Sergio.pdf
accesso aperto
Dimensione
8.04 MB
Formato
Adobe PDF
|
8.04 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/114302