Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Biochemical Prostate Cancer Recurrence Prediction: Thinking Fast & Slow

Suhang You, Sanyukta Adap, Siddhesh Thakur, Bhakti Baheti, Spyridon Bakas

TL;DR

This work proposed to leverage multiple instance learning through a two-stage ``thinking fast \&slow'' strategy for the time to recurrence (TTR) prediction.

Abstract

Time to biochemical recurrence in prostate cancer is essential for prognostic monitoring of the progression of patients after prostatectomy, which assesses the efficacy of the surgery. In this work, we proposed to leverage multiple instance learning through a two-stage ``thinking fast \& slow'' strategy for the time to recurrence (TTR) prediction. The first (``thinking fast'') stage finds the most relevant WSI area for biochemical recurrence and the second (``thinking slow'') stage leverages higher resolution patches to predict TTR. Our approach reveals a mean C-index ($Ci$) of 0.733 ($θ=0.059$) on our internal validation and $Ci=0.603$ on the LEOPARD challenge validation set. Post hoc attention visualization shows that the most attentive area contributes to the TTR prediction.

Biochemical Prostate Cancer Recurrence Prediction: Thinking Fast & Slow

TL;DR

This work proposed to leverage multiple instance learning through a two-stage ``thinking fast \&slow'' strategy for the time to recurrence (TTR) prediction.

Abstract

Time to biochemical recurrence in prostate cancer is essential for prognostic monitoring of the progression of patients after prostatectomy, which assesses the efficacy of the surgery. In this work, we proposed to leverage multiple instance learning through a two-stage ``thinking fast \& slow'' strategy for the time to recurrence (TTR) prediction. The first (``thinking fast'') stage finds the most relevant WSI area for biochemical recurrence and the second (``thinking slow'') stage leverages higher resolution patches to predict TTR. Our approach reveals a mean C-index () of 0.733 () on our internal validation and on the LEOPARD challenge validation set. Post hoc attention visualization shows that the most attentive area contributes to the TTR prediction.
Paper Structure (10 sections, 4 equations, 3 figures)

This paper contains 10 sections, 4 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Material
  3. Methods
  4. Thinking Fast: MIL Classification at Low Resolution
  5. Thinking Slow: MIL Regression at High Resolution
  6. Model Training, Evaluation & Selection
  7. Results
  8. Interpretability
  9. Discussion
  10. Code Link

Figures (3)

  • Figure 1: Illustration of our two-stage "Thinking fast & slow" approach. During "thinking fast", a patch mask is generated to rapidly localize relevant WSI patches. During "thinking slow" stage, top k and attention pooling are used for WSI TTR prediction.
  • Figure 2: Ablation results for model selection. (A) and (C) show the comparison among different top k settings at different percentage of WSI ($m$). (B) and (D) compare the best parameter setting across MAD-MIL, AC-MIL, and our MIL method. y-axis in (A) and (B) represents mean $Ci$ values evaluated on the outer hold-out set and y-axis in (C) and (D) represents $\sigma_{Ci}$. The x-axis represents the percentage of patches used for the "thinking slow" stage.
  • Figure 3: One example shows interpretability for our model. Both images are overlays of the attention heat map on the tissue. The image on the right is a zoom-in for the selected area on the left. Higher attention scores are in more red-covered area.