-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathshape.html
More file actions
207 lines (195 loc) · 10.8 KB
/
Copy pathshape.html
File metadata and controls
207 lines (195 loc) · 10.8 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<title>SHAPE — Segmentation, Hybrid, And Point‑based Evaluation for LV Border Extraction in Echocardiography (A4C)</title>
<meta name="viewport" content="width=device-width, initial-scale=1" />
<style>
:root{
--bg:#f8fafc;
--fg:#0f172a;
--muted:#475569;
--accent:#2563eb;
--card:#ffffff;
--border:#e2e8f0;
}
html,body{margin:0;padding:0;background:var(--bg);color:var(--fg);font-family:Inter,system-ui,-apple-system,Segoe UI,Roboto,Ubuntu,Cantarell,"Helvetica Neue",Arial,"Noto Sans","Apple Color Emoji","Segoe UI Emoji","Segoe UI Symbol";line-height:1.55}
.page{max-width:980px;margin:48px auto;padding:0 24px 64px}
header{background:linear-gradient(135deg,#eff6ff,#ffffff);border:1px solid var(--border);padding:28px 24px;border-radius:16px}
h1{margin:0 0 8px 0;font-size:clamp(26px,3.6vw,36px);line-height:1.2}
.subtitle{color:var(--muted);font-size:clamp(14px,2.2vw,16px)}
.badge{display:inline-block;background:var(--accent);color:#fff;border-radius:999px;padding:6px 12px;font-weight:600;letter-spacing:.2px;margin:8px 0}
nav{margin:28px 0 8px}
nav a{display:inline-block;margin:6px 10px 0 0;color:var(--accent);text-decoration:none;border:1px solid var(--border);background:var(--card);padding:6px 10px;border-radius:8px}
section{background:var(--card);border:1px solid var(--border);border-radius:16px;padding:20px;margin:16px 0}
h2{margin-top:0}
h3{margin:18px 0 8px}
ul{margin:8px 0 8px 20px}
.grid{display:grid;gap:12px}
@media(min-width:820px){.grid-2{grid-template-columns:1fr 1fr}}
.muted{color:var(--muted)}
.code{font-family:ui-monospace,SFMono-Regular,Menlo,Monaco,Consolas,"Liberation Mono","Courier New",monospace;background:#0b1220;color:#e5e7eb;border-radius:10px;padding:12px;overflow:auto}
.callout{border-left:4px solid var(--accent);background:#f1f5ff;padding:12px 14px;border-radius:8px}
.kpi{display:flex;gap:16px;flex-wrap:wrap}
.kpi div{flex:1 1 220px;background:#f9fafb;border:1px solid var(--border);border-radius:12px;padding:12px}
table{width:100%;border-collapse:separate;border-spacing:0;margin:8px 0}
th,td{border-bottom:1px solid var(--border);padding:10px 8px;text-align:left;vertical-align:top}
thead th{background:#f8fafc;font-weight:600}
.footnotes{font-size:0.92rem;color:var(--muted)}
.small{font-size:0.95rem}
.btn{display:inline-block;margin-top:8px;color:#fff;background:var(--accent);padding:8px 12px;border-radius:8px;text-decoration:none}
@media print{
body{background:#fff}
header,section{break-inside:avoid}
nav{display:none}
}
</style>
</head>
<body>
<div class="page">
<header>
<span class="badge">SHAPE</span>
<h1>Segmentation, Hybrid, And Point‑based Evaluation for LV Border Extraction in Echocardiography (A4C)</h1>
<p class="subtitle">MSc Project Proposal • Comparative analysis of deep learning strategies for left ventricular (LV) border extraction in the Apical Four‑Chamber view</p>
<nav>
<a href="#summary">Summary</a>
<a href="#aim">Aim & Questions</a>
<a href="#methodology">Methodology</a>
<a href="#data">Data & Annotations</a>
<a href="#evaluation">Evaluation</a>
</nav>
</header>
<section id="summary">
<h2>1) Introduction & Background</h2>
<p>
Accurate and reproducible delineation of the LV endocardial border in two‑dimensional echocardiography is essential for
deriving LV area, volumes (EDV/ESV), LVID<sub>d/s</sub>, and ultimately LVEF. Manual tracing is labour‑intensive and subject to
inter‑observer variability, while echo images often suffer from low contrast and speckle noise. Deep learning offers routes to
automation, typically via (i) pixel‑wise <strong>segmentation</strong>, (ii) <strong>point‑based</strong> landmark detection (apex and annulus points),
or (iii) <strong>hybrid multitask</strong> models combining both. This project focuses on the Apical Four‑Chamber (A4C) view, a primary
plane for biplane volumetry and routine clinical reporting.
</p>
<div class="kpi">
<div><strong>Clinical target</strong><br/><span class="muted">Reliable A4C borders for area, LVID, and single‑plane Simpson volumes</span></div>
<div><strong>Comparators</strong><br/><span class="muted">Segmentation vs Point‑based vs Hybrid (mask+keypoints)</span></div>
<div><strong>Outcome</strong><br/><span class="muted">Which strategy yields the most clinically useful and robust borders?</span></div>
</div>
</section>
<section id="aim">
<h2>2) Aim & Research Questions</h2>
<p><strong>Aim.</strong> Compare segmentation‑only, point‑based landmark, and hybrid multitask approaches for LV border extraction in A4C, evaluating clinical utility, robustness on low‑quality frames, and generalisation across imaging conditions.</p>
<h3>Key questions</h3>
<ul>
<li><strong>Clinical utility:</strong> Which approach yields the most accurate borders for LV area, LVID<sub>d/s</sub>, and A4C‑based volumes?</li>
<li><strong>Robustness:</strong> Do explicit landmarks (apex, septal/lateral annulus) improve performance on low‑quality frames?</li>
<li><strong>Generalisation:</strong> Does multitask supervision (mask+keypoints) generalise better across vendors/sites?</li>
</ul>
</section>
<section id="methodology">
<h2>3) Methodology (A4C‑only)</h2>
<div class="grid grid-2">
<div>
<h3>Track A — Segmentation</h3>
<ul>
<li><strong>Model:</strong> U‑Net/UNet++/nnU‑Net (PyTorch/MONAI).</li>
<li><strong>Loss:</strong> Dice + Cross‑Entropy + <em>contour‑aware term</em> (e.g., distance transform / Chamfer) to sharpen borders.</li>
<li><strong>Output:</strong> LV mask → contour via marching squares / border tracing.</li>
</ul>
</div>
<div>
<h3>Track B — Point‑based (Landmarks)</h3>
<ul>
<li><strong>Model:</strong> Heatmap regression (e.g., HRNet / stacked hourglass).</li>
<li><strong>Targets:</strong> Apex, septal annulus, lateral annulus.</li>
<li><strong>Post‑proc:</strong> Spline or snake initialised by landmarks to form a smooth border.</li>
</ul>
</div>
</div>
<div class="grid grid-2">
<div>
<h3>Track C — Hybrid Multitask</h3>
<ul>
<li><strong>Backbone:</strong> Shared encoder; two heads (mask & heatmaps).</li>
<li><strong>Composite loss:</strong>
<span class="small">
λ<sub>1</sub>(Dice+CE) + λ<sub>2</sub>MSE<sub>heatmaps</sub> + λ<sub>3</sub> (Chamfer/Hausdorff on final contour)
+ optional landmark‑contour consistency.
</span>
</li>
<li><strong>Fusion:</strong> Combine predicted mask + landmarks (graph shortest‑path or snake) → final contour.</li>
</ul>
</div>
<div>
<h3>Pre‑/Post‑processing</h3>
<ul>
<li>Standardise pixel spacing; crop ROI around LV; intensity normalisation.</li>
<li>Augmentations: rotation (±10°), scale, gamma, mild elastic, compression artefacts.</li>
<li>ED/ES selection provided or detected via area curve from coarse segmentation.</li>
</ul>
</div>
</div>
<div class="callout">
<strong>Note on clinical plausibility.</strong> Landmark constraints (apex & annulus line) can regularise contours and reduce anatomically implausible shapes on noisy frames.
</div>
</section>
<section id="data">
<h2>4) Data & Annotation Strategy (General)</h2>
<p>
The study will use <strong>A4C cine frames</strong> and corresponding ED/ES images sourced from <em>either</em> open datasets (subject to licence)
<em>or</em> institutional repositories under an approved governance protocol. Exact sources are flexible to local availability; the experimental
design does not depend on a specific dataset brand.
</p>
<h3>Ground truth (two label types)</h3>
<ul>
<li><strong>Masks:</strong> Expert endocardial tracings (ED & ES).</li>
<li><strong>Landmarks:</strong> Apex; septal and lateral annulus points (mitral hinge line).</li>
</ul>
<h3>When only one label type is available</h3>
<ul>
<li><strong>Derive landmarks from masks:</strong> (i) estimate annulus line by detecting valve‑hinge boundary intersections;
(ii) define apex as farthest point from the annulus line along the contour; (iii) optional curvature checks.</li>
<li><strong>Derive weak masks from landmarks:</strong> fit a spline through landmarks, rasterise to a thin mask, optionally dilate (2–3 px)
for supervision in low‑label settings.</li>
</ul>
<h3>Quality control</h3>
<ul>
<li>Dual annotation on 10–15% to estimate inter‑observer variability (used as non‑inferiority margin).</li>
<li>Manual validation of automatically derived landmarks on a stratified subset.</li>
</ul>
</section>
<section id="evaluation">
<h2>5) Evaluation Plan</h2>
<div class="grid grid-2">
<div>
<h3>Contour quality</h3>
<ul>
<li>Dice, IoU (overlap)</li>
<li>Hausdorff95, mean surface distance, point‑to‑curve error (mm)</li>
<li>Landmark error (mm) at apex/annuli</li>
</ul>
<h3>Robustness & generalisation</h3>
<ul>
<li>Stratify by image quality, vendor/site (if available), and stress tests (noise, blur, rotation)</li>
<li>Uncertainty: test‑time augmentation / MC‑dropout; calibration (ECE/AUCE)</li>
</ul>
</div>
<div>
<h3>Clinical accuracy (A4C)</h3>
<ul>
<li>LV area (ED/ES), long‑axis length (annulus–apex), single‑plane Simpson EDV/ESV</li>
<li>Agreement vs clinician: Bland–Altman, MAE/RMSE, Pearson/Spearman, 95% LoA coverage</li>
<li>Failure rate on low‑quality frames; time‑per‑frame</li>
</ul>
<h3>Statistics</h3>
<ul>
<li>Paired tests across methods (Wilcoxon/t‑test); effect sizes & 95% CI</li>
<li>Non‑inferiority vs inter‑observer variability</li>
</ul>
</div>
</div>
<h3>Primary endpoint</h3>
<p>Lowest EDV/ESV MAE (A4C single‑plane Simpson) and narrowest limits of agreement, or equivalent performance with fewer labelled masks (data‑efficiency advantage).</p>
</section>
</div>
</body>
</html>