File size: 13,247 Bytes
1d197a4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
#!/usr/bin/env python3
"""Generate figures and numeric summaries for the project fine-tuning memo."""

from __future__ import annotations

import json
import math
from pathlib import Path

import matplotlib.pyplot as plt
import numpy as np

ROOT = Path(__file__).resolve().parents[2]
ASSET_DIR = ROOT / "docs" / "memo_assets"
ASSET_DIR.mkdir(parents=True, exist_ok=True)


# ---------------------------- Helpers ----------------------------

def read_json(path: Path) -> dict:
    with path.open("r", encoding="utf-8") as f:
        return json.load(f)


def read_json_first(paths: list[Path], required: bool = True) -> dict:
    for path in paths:
        if path.exists():
            return read_json(path)
    if required:
        raise FileNotFoundError(f"None of the candidate files exist: {paths}")
    return {}


def wilson_interval(successes: int, n: int, z: float = 1.96) -> tuple[float, float]:
    if n <= 0:
        return float("nan"), float("nan")
    p = successes / n
    denom = 1.0 + (z * z / n)
    center = (p + (z * z / (2 * n))) / denom
    half = (z / denom) * math.sqrt((p * (1 - p) / n) + (z * z / (4 * n * n)))
    return max(0.0, center - half), min(1.0, center + half)


def ensure_style():
    plt.rcParams.update(
        {
            "font.size": 10,
            "axes.grid": True,
            "grid.alpha": 0.22,
            "figure.facecolor": "white",
            "axes.facecolor": "#fafafa",
        }
    )


# ---------------------------- Load inputs ----------------------------

multitask_summary = read_json_first(
    [
        ROOT / "models" / "multitask" / "multitask_summary.json",
        ROOT / "output" / "training_outputs" / "multitask_summary.json",
    ],
    required=False,
)
multitask_test = read_json_first(
    [
        ROOT / "output" / "multitask_test_inference.json",
        ROOT / "output" / "training_outputs" / "multitask_test_inference.json",
    ]
)
lumen_summary = read_json_first(
    [
        ROOT / "models" / "standalone" / "lumen" / "finetune_summary.json",
        ROOT / "output" / "training_outputs" / "lumen_finetune_summary.json",
    ]
)
standalone_test = read_json_first(
    [
        ROOT / "output" / "bifurcation_classifier" / "test_inference.json",
        ROOT / "output" / "training_outputs" / "bifurcation_classifier" / "test_inference.json",
    ]
)
threshold_payload = read_json_first(
    [
        ROOT / "models" / "multitask" / "threshold.json",
        ROOT / "models" / "standalone" / "bifurcation" / "threshold.json",
        ROOT / "models" / "bifurcation" / "threshold.json",
        ROOT / "output" / "bifurcation_classifier" / "threshold.json",
    ]
)
dataset_balance = read_json_first(
    [
        ROOT / "output" / "dataset_balance" / "dataset_balance_summary.json",
        ROOT / "docs" / "memo_assets" / "dataset_balance_summary.json",
    ]
)


# ---------------------------- Figure 1: Training curves (multitask) ----------------------------

ensure_style()
history = multitask_summary.get("history", [])
if history:
    epochs = [int(r["epoch"]) for r in history]
    train_total = [float(r.get("total_loss", np.nan)) for r in history]
    val_total = [float(r.get("val_total_loss", np.nan)) for r in history]
    val_dice = [float(r.get("val_seg_dice", np.nan)) for r in history]
    val_iou = [float(r.get("val_seg_iou", np.nan)) for r in history]
    val_f1 = [float(r.get("val_cls_f1", np.nan)) for r in history]
    val_auc = [float(r.get("val_cls_auc", np.nan)) for r in history]

    fig, axes = plt.subplots(2, 2, figsize=(12, 8))

    axes[0, 0].plot(epochs, train_total, label="train total loss", color="#1f77b4")
    axes[0, 0].plot(epochs, val_total, label="val total loss", color="#d62728")
    axes[0, 0].set_title("Multitask Total Loss")
    axes[0, 0].set_xlabel("Epoch")
    axes[0, 0].set_ylabel("Loss")
    axes[0, 0].legend()

    axes[0, 1].plot(epochs, val_dice, label="val seg dice", color="#2ca02c")
    axes[0, 1].plot(epochs, val_iou, label="val seg IoU", color="#17becf")
    axes[0, 1].set_title("Validation Segmentation Metrics")
    axes[0, 1].set_xlabel("Epoch")
    axes[0, 1].set_ylabel("Score")
    axes[0, 1].set_ylim(0.75, 0.97)
    axes[0, 1].legend()

    axes[1, 0].plot(epochs, val_f1, label="val cls F1", color="#9467bd")
    axes[1, 0].plot(epochs, val_auc, label="val cls AUC", color="#ff7f0e")
    axes[1, 0].set_title("Validation Classification Metrics")
    axes[1, 0].set_xlabel("Epoch")
    axes[1, 0].set_ylabel("Score")
    axes[1, 0].set_ylim(0.0, 1.0)
    axes[1, 0].legend()

    # Stability plot: train seg/cls losses (if available)
    train_seg = [float(r.get("seg_loss", np.nan)) for r in history]
    train_cls = [float(r.get("cls_loss", np.nan)) for r in history]
    axes[1, 1].plot(epochs, train_seg, label="train seg loss", color="#8c564b")
    axes[1, 1].plot(epochs, train_cls, label="train cls loss", color="#e377c2")
    axes[1, 1].set_title("Train Task Loss Components")
    axes[1, 1].set_xlabel("Epoch")
    axes[1, 1].set_ylabel("Loss")
    axes[1, 1].legend()

    fig.suptitle("Multitask Fine-tuning Dynamics", fontsize=13)
    fig.tight_layout()
    fig.savefig(ASSET_DIR / "multitask_training_dynamics.png", dpi=220)
    plt.close(fig)


# ---------------------------- Figure 2: Lumen fine-tune curves ----------------------------

lh = lumen_summary.get("history", [])
le = [int(float(r.get("epoch", i + 1))) for i, r in enumerate(lh)]
tr_loss = [float(r.get("train_loss", np.nan)) for r in lh]
val_loss = [float(r.get("val_loss", np.nan)) for r in lh]
val_dice_l = [float(r.get("val_dice", np.nan)) for r in lh]
val_iou_l = [float(r.get("val_iou", np.nan)) for r in lh]

fig, axes = plt.subplots(1, 2, figsize=(11, 4.2))
axes[0].plot(le, tr_loss, label="train loss", color="#1f77b4")
axes[0].plot(le, val_loss, label="val loss", color="#d62728")
axes[0].set_title("Lumen Fine-tune Loss")
axes[0].set_xlabel("Epoch")
axes[0].set_ylabel("Loss")
axes[0].legend()

axes[1].plot(le, val_dice_l, label="val dice", color="#2ca02c")
axes[1].plot(le, val_iou_l, label="val IoU", color="#17becf")
axes[1].set_title("Lumen Fine-tune Validation Metrics")
axes[1].set_xlabel("Epoch")
axes[1].set_ylabel("Score")
axes[1].set_ylim(0.82, 0.97)
axes[1].legend()

fig.tight_layout()
fig.savefig(ASSET_DIR / "lumen_finetune_dynamics.png", dpi=220)
plt.close(fig)


# ---------------------------- Figure 3: Standalone threshold sweep ----------------------------

sweep = standalone_test.get("threshold_selection", {}).get("val_threshold_sweep", [])
if sweep:
    ths = np.array([float(r["threshold"]) for r in sweep], dtype=float)
    f1s = np.array([float(r.get("f1", np.nan)) for r in sweep], dtype=float)
    recs = np.array([float(r.get("recall", np.nan)) for r in sweep], dtype=float)
    precs = np.array([float(r.get("precision", np.nan)) for r in sweep], dtype=float)

    best_t = float(standalone_test.get("selected_threshold", ths[int(np.nanargmax(f1s))]))
    best_i = int(np.argmin(np.abs(ths - best_t)))

    fig, ax = plt.subplots(figsize=(8, 4.8))
    ax.plot(ths, f1s, label="F1", color="#1f77b4", linewidth=2)
    ax.plot(ths, recs, label="Recall", color="#2ca02c", alpha=0.75)
    ax.plot(ths, precs, label="Precision", color="#d62728", alpha=0.75)
    ax.scatter([ths[best_i]], [f1s[best_i]], color="black", s=60, zorder=3, label=f"Selected t={ths[best_i]:.3f}")
    ax.axvline(ths[best_i], linestyle="--", color="black", linewidth=1)
    ax.set_title("Validation Threshold Sweep (Standalone Bifurcation Model)")
    ax.set_xlabel("Threshold")
    ax.set_ylabel("Score")
    ax.set_ylim(0.0, 1.0)
    ax.legend(loc="best")
    fig.tight_layout()
    fig.savefig(ASSET_DIR / "standalone_threshold_sweep.png", dpi=220)
    plt.close(fig)


# ---------------------------- Figure 4: Probability distributions ----------------------------

pred_rows = standalone_test.get("predictions", [])
if pred_rows:
    y_true = np.array([int(r["label"]) for r in pred_rows], dtype=int)
    y_prob = np.array([float(r["probability"]) for r in pred_rows], dtype=float)
    th = float(standalone_test.get("selected_threshold", 0.5))

    fig, ax = plt.subplots(figsize=(8, 4.6))
    ax.hist(y_prob[y_true == 0], bins=24, alpha=0.6, label="True non-bifurcation", color="#1f77b4")
    ax.hist(y_prob[y_true == 1], bins=24, alpha=0.6, label="True bifurcation", color="#d62728")
    ax.axvline(th, color="black", linestyle="--", linewidth=1.3, label=f"Operating threshold={th:.3f}")
    ax.set_title("Predicted Probability Distributions (Standalone Test Set)")
    ax.set_xlabel("Predicted bifurcation probability")
    ax.set_ylabel("Count")
    ax.legend()
    fig.tight_layout()
    fig.savefig(ASSET_DIR / "standalone_probability_hist.png", dpi=220)
    plt.close(fig)


# ---------------------------- Figure 5: Pipeline diagram ----------------------------

fig, ax = plt.subplots(figsize=(13, 5.8))
ax.axis("off")

# Draw boxes manually
boxes = {
    "data": (0.03, 0.62, 0.2, 0.25, "Frame-bank JSONL + DICOM\n(train/val/test split)"),
    "pre": (0.28, 0.62, 0.2, 0.25, "Preprocess\n(center black circle,\n3-channel replication)"),
    "base": (0.53, 0.62, 0.2, 0.25, "Pretrained lumen base\n(SavedModel logits)"),
    "seg": (0.78, 0.72, 0.19, 0.14, "Segmentation branch\n(BCE + Dice)") ,
    "cls": (0.78, 0.52, 0.19, 0.14, "Classification head\n(GAP + Dense + Sigmoid)") ,
    "loss": (0.53, 0.18, 0.26, 0.22, "Total loss\n= w_seg * seg_loss\n+ w_cls * cls_loss\n(mask seg loss by has_mask)"),
    "eval": (0.84, 0.18, 0.13, 0.22, "Val threshold sweep\n→ threshold.json\n→ runtime inference"),
}

for _, (x, y, w, h, txt) in boxes.items():
    rect = plt.Rectangle((x, y), w, h, facecolor="#f7f7f7", edgecolor="#333333", linewidth=1.2)
    ax.add_patch(rect)
    ax.text(x + w / 2, y + h / 2, txt, ha="center", va="center", fontsize=10)


def arrow(x0, y0, x1, y1):
    ax.annotate("", xy=(x1, y1), xytext=(x0, y0), arrowprops=dict(arrowstyle="->", linewidth=1.4, color="#333333"))

arrow(0.23, 0.745, 0.28, 0.745)
arrow(0.48, 0.745, 0.53, 0.745)
arrow(0.73, 0.745, 0.78, 0.79)
arrow(0.73, 0.71, 0.78, 0.59)
arrow(0.875, 0.72, 0.66, 0.40)
arrow(0.875, 0.52, 0.66, 0.40)
arrow(0.79, 0.29, 0.84, 0.29)

ax.set_title("Training and Inference Design: Pretrained Base + Multitask Head", fontsize=13)
fig.tight_layout()
fig.savefig(ASSET_DIR / "multitask_pipeline_diagram.png", dpi=220)
plt.close(fig)


# ---------------------------- Figure 6: Metric comparison ----------------------------

mt = multitask_test["bifurcation_metrics"]
ms = multitask_test["segmentation_metrics"]
lumen_test = lumen_summary["final_test_metrics"]

labels = ["Lumen IoU", "Lumen Dice", "Bif Acc", "Bif F1", "Bif AUC"]
values = [
    float(ms.get("seg_iou", np.nan)),
    float(ms.get("seg_dice", np.nan)),
    float(mt.get("cls_accuracy", np.nan)),
    float(mt.get("cls_f1", np.nan)),
    float(mt.get("cls_auc", np.nan)),
]

fig, ax = plt.subplots(figsize=(8.2, 4.4))
bars = ax.bar(labels, values, color=["#17becf", "#2ca02c", "#1f77b4", "#9467bd", "#ff7f0e"])
ax.set_ylim(0.0, 1.0)
ax.set_ylabel("Score")
ax.set_title("Multitask Test Metrics Snapshot")
for b, v in zip(bars, values):
    ax.text(b.get_x() + b.get_width() / 2, v + 0.015, f"{v:.3f}", ha="center", va="bottom", fontsize=9)
fig.tight_layout()
fig.savefig(ASSET_DIR / "multitask_test_metric_snapshot.png", dpi=220)
plt.close(fig)


# ---------------------------- Summary JSON ----------------------------

cls = multitask_test["bifurcation_metrics"]
seg = multitask_test["segmentation_metrics"]
n_cls = int(cls["tp"] + cls["fp"] + cls["fn"] + cls["tn"])
acc_ci = wilson_interval(int(cls["tp"] + cls["tn"]), n_cls)
rec_ci = wilson_interval(int(cls["tp"]), int(cls["tp"] + cls["fn"]))
prec_ci = wilson_interval(int(cls["tp"]), int(cls["tp"] + cls["fp"]))

best_epoch = None
if history:
    best_epoch = min(history, key=lambda r: float(r.get("val_total_loss", np.inf))).get("epoch")

summary = {
    "data": {
        "split_counts": dataset_balance.get("counts", {}),
        "bifurcation_positive_rate": dataset_balance.get("bifurcation_positive_rate", {}),
        "lumen_coverage_rate": dataset_balance.get("lumen_coverage_rate", {}),
    },
    "multitask": {
        "epochs_completed": len(history),
        "best_epoch_by_val_total_loss": best_epoch,
        "best_val_total_loss": multitask_summary.get("best_val_total_loss"),
        "test_seg_metrics": seg,
        "test_cls_metrics": cls,
        "cls_accuracy_wilson95": acc_ci,
        "cls_precision_wilson95": prec_ci,
        "cls_recall_wilson95": rec_ci,
        "threshold": multitask_test.get("threshold_info", {}).get("selected_threshold"),
    },
    "lumen_finetune": {
        "num_samples": lumen_summary.get("num_samples"),
        "test_metrics": lumen_summary.get("final_test_metrics", {}),
    },
    "standalone_bifurcation": {
        "selected_threshold": standalone_test.get("selected_threshold"),
        "test_metrics": standalone_test.get("metrics", {}),
    },
    "runtime_threshold": threshold_payload.get("selected_threshold"),
}

with (ASSET_DIR / "memo_summary.json").open("w", encoding="utf-8") as f:
    json.dump(summary, f, indent=2)

print("Saved assets to", ASSET_DIR)
for p in sorted(ASSET_DIR.glob("*")):
    print(p.name)